security: Enhance Hugging Face model loading security - resolves #136

BREAKING CHANGE: trust_remote_code now defaults to False for security

- Set trust_remote_code=False by default in HFChat class
- Add explicit trust_remote_code parameter to HFChat.__init__()
- Add security warning when trust_remote_code=True is used
- Update get_llm() function to support trust_remote_code parameter
- Update benchmark utilities (load_hf_model, load_vllm_model, load_qwen_vl_model)
- Add comprehensive documentation for security implications

Security Benefits:
- Prevents arbitrary code execution from compromised model repositories
- Requires explicit opt-in for models that need remote code execution
- Shows clear warnings when security is reduced
- Follows security-by-default principle

Migration Guide:
- Most users: No changes needed (more secure by default)
- Users with models requiring remote code: Add trust_remote_code=True explicitly
- Config users: Add 'trust_remote_code': true to LLM config if needed

Fixes #136

.gitattributes

@@ -1 +0,0 @@
-paper_plot/data/big_graph_degree_data.npz filter=lfs diff=lfs merge=lfs -text

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,50 @@
+name: Bug Report
+description: Report a bug in LEANN
+labels: ["bug"]
+
+body:
+  - type: textarea
+    id: description
+    attributes:
+      label: What happened?
+      description: A clear description of the bug
+    validations:
+      required: true
+
+  - type: textarea
+    id: reproduce
+    attributes:
+      label: How to reproduce
+      placeholder: |
+        1. Install with...
+        2. Run command...
+        3. See error
+    validations:
+      required: true
+
+  - type: textarea
+    id: error
+    attributes:
+      label: Error message
+      description: Paste any error messages
+      render: shell
+
+  - type: input
+    id: version
+    attributes:
+      label: LEANN Version
+      placeholder: "0.1.0"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: os
+    attributes:
+      label: Operating System
+      options:
+        - macOS
+        - Linux
+        - Windows
+        - Docker
+    validations:
+      required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,8 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Documentation
+    url: https://github.com/LEANN-RAG/LEANN-RAG/tree/main/docs
+    about: Read the docs first
+  - name: Discussions
+    url: https://github.com/LEANN-RAG/LEANN-RAG/discussions
+    about: Ask questions and share ideas

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,27 @@
+name: Feature Request
+description: Suggest a new feature for LEANN
+labels: ["enhancement"]
+
+body:
+  - type: textarea
+    id: problem
+    attributes:
+      label: What problem does this solve?
+      description: Describe the problem or need
+    validations:
+      required: true
+
+  - type: textarea
+    id: solution
+    attributes:
+      label: Proposed solution
+      description: How would you like this to work?
+    validations:
+      required: true
+
+  - type: textarea
+    id: example
+    attributes:
+      label: Example usage
+      description: Show how the API might look
+      render: python

.github/pull_request_template.md

@@ -0,0 +1,13 @@
+## What does this PR do?
+
+<!-- Brief description of your changes -->
+
+## Related Issues
+
+Fixes #
+
+## Checklist
+
+- [ ] Tests pass (`uv run pytest`)
+- [ ] Code formatted (`ruff format` and `ruff check`)
+- [ ] Pre-commit hooks pass (`pre-commit run --all-files`)

.github/workflows/build-and-publish.yml

@@ -1,256 +1,12 @@
-name: Build and Publish to PyPI
+name: CI
 
 on:
-  release:
-    types: [published]
   push:
-    tags:
-      - 'v*'
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
   workflow_dispatch:
-    inputs:
-      publish:
-        description: 'Publish to PyPI'
-        required: true
-        default: 'false'
-        type: choice
-        options:
-          - 'false'
-          - 'test'
-          - 'prod'
 
 jobs:
-  # Build pure Python package: leann-core
-  build-core:
-    name: Build leann-core
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-        
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-          
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-          
-      - name: Install build dependencies
-        run: |
-          uv pip install --system build twine
-          
-      - name: Build package
-        run: |
-          cd packages/leann-core
-          uv build
-          
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: leann-core-dist
-          path: packages/leann-core/dist/
-
-  # Build binary package: leann-backend-hnsw (default backend)
-  build-hnsw:
-    name: Build leann-backend-hnsw
-    strategy:
-      matrix:
-        os: [ubuntu-latest, macos-latest]
-        python-version: ['3.9', '3.10', '3.11', '3.12']
-    runs-on: ${{ matrix.os }}
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-          
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python-version }}
-          
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-          
-      - name: Install system dependencies (Ubuntu)
-        if: runner.os == 'Linux'
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y libomp-dev libboost-all-dev libzmq3-dev \
-            pkg-config libopenblas-dev patchelf
-          
-      - name: Install system dependencies (macOS)
-        if: runner.os == 'macOS'
-        run: |
-          brew install libomp boost zeromq
-          
-      - name: Install build dependencies
-        run: |
-          uv pip install --system scikit-build-core numpy swig
-          uv pip install --system auditwheel delocate
-          
-      - name: Build wheel
-        run: |
-          cd packages/leann-backend-hnsw
-          uv build --wheel
-          
-      - name: Repair wheel (Linux)
-        if: runner.os == 'Linux'
-        run: |
-          cd packages/leann-backend-hnsw
-          auditwheel repair dist/*.whl -w dist_repaired
-          rm -rf dist
-          mv dist_repaired dist
-          
-      - name: Repair wheel (macOS)
-        if: runner.os == 'macOS'
-        run: |
-          cd packages/leann-backend-hnsw
-          delocate-wheel -w dist_repaired -v dist/*.whl
-          rm -rf dist
-          mv dist_repaired dist
-          
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: hnsw-${{ matrix.os }}-py${{ matrix.python-version }}
-          path: packages/leann-backend-hnsw/dist/
-
-  # Build binary package: leann-backend-diskann (multi-platform)
-  build-diskann:
-    name: Build leann-backend-diskann
-    strategy:
-      matrix:
-        os: [ubuntu-latest, macos-latest]
-        python-version: ['3.9', '3.10', '3.11', '3.12']
-    runs-on: ${{ matrix.os }}
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-          
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python-version }}
-          
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-          
-      - name: Install system dependencies (Ubuntu)
-        if: runner.os == 'Linux'
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y libomp-dev libboost-all-dev libaio-dev libzmq3-dev \
-            protobuf-compiler libprotobuf-dev libabsl-dev patchelf
-          
-          # Install Intel MKL using Intel's installer
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/79153e0f-74d7-45af-b8c2-258941adf58a/intel-onemkl-2025.0.0.940.sh
-          sudo sh intel-onemkl-2025.0.0.940.sh -a --components intel.oneapi.lin.mkl.devel --action install --eula accept -s
-          source /opt/intel/oneapi/setvars.sh
-          echo "MKLROOT=/opt/intel/oneapi/mkl/latest" >> $GITHUB_ENV
-          echo "LD_LIBRARY_PATH=/opt/intel/oneapi/mkl/latest/lib/intel64:$LD_LIBRARY_PATH" >> $GITHUB_ENV
-          
-      - name: Install system dependencies (macOS)
-        if: runner.os == 'macOS'
-        run: |
-          brew install libomp boost zeromq protobuf
-          # MKL is not available on Homebrew, but DiskANN can work without it
-          
-      - name: Install build dependencies
-        run: |
-          uv pip install --system scikit-build-core numpy Cython pybind11
-          if [[ "$RUNNER_OS" == "Linux" ]]; then
-            uv pip install --system auditwheel
-          else
-            uv pip install --system delocate
-          fi
-          
-      - name: Build wheel
-        run: |
-          cd packages/leann-backend-diskann
-          uv build --wheel
-          
-      - name: Repair wheel (Linux)
-        if: runner.os == 'Linux'
-        run: |
-          cd packages/leann-backend-diskann
-          auditwheel repair dist/*.whl -w dist_repaired
-          rm -rf dist
-          mv dist_repaired dist
-          
-      - name: Repair wheel (macOS)
-        if: runner.os == 'macOS'
-        run: |
-          cd packages/leann-backend-diskann
-          delocate-wheel -w dist_repaired -v dist/*.whl
-          rm -rf dist
-          mv dist_repaired dist
-          
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: diskann-${{ matrix.os }}-py${{ matrix.python-version }}
-          path: packages/leann-backend-diskann/dist/
-
-  # Build meta-package: leann (build last)
-  build-meta:
-    name: Build leann meta-package
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-        
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-          
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-          
-      - name: Install build dependencies
-        run: |
-          uv pip install --system build
-          
-      - name: Build package
-        run: |
-          cd packages/leann
-          uv build
-          
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: leann-meta-dist
-          path: packages/leann/dist/
-
-  # Publish to PyPI
-  publish:
-    name: Publish to PyPI
-    needs: [build-core, build-hnsw, build-diskann, build-meta]
-    runs-on: ubuntu-latest
-    if: github.event_name == 'release' || github.event.inputs.publish != 'false'
-    
-    steps:
-      - name: Download all artifacts
-        uses: actions/download-artifact@v4
-        with:
-          path: dist
-          
-      - name: Flatten directory structure
-        run: |
-          mkdir -p all_wheels
-          find dist -name "*.whl" -exec cp {} all_wheels/ \;
-          find dist -name "*.tar.gz" -exec cp {} all_wheels/ \;
-          
-      - name: Publish to Test PyPI
-        if: github.event.inputs.publish == 'test' || github.event_name == 'workflow_dispatch'
-        uses: pypa/gh-action-pypi-publish@release/v1
-        with:
-          password: ${{ secrets.TEST_PYPI_API_TOKEN }}
-          repository-url: https://test.pypi.org/legacy/
-          packages-dir: all_wheels/
-          
-      - name: Publish to PyPI
-        if: github.event_name == 'release' || github.event.inputs.publish == 'prod'
-        uses: pypa/gh-action-pypi-publish@release/v1
-        with:
-          password: ${{ secrets.PYPI_API_TOKEN }}
-          packages-dir: all_wheels/ 
+  build:
+    uses: ./.github/workflows/build-reusable.yml

.github/workflows/build-reusable.yml

@@ -0,0 +1,451 @@
+name: Reusable Build
+
+on:
+  workflow_call:
+    inputs:
+      ref:
+        description: 'Git ref to build'
+        required: false
+        type: string
+        default: ''
+
+jobs:
+  lint:
+    name: Lint and Format Check
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+          submodules: recursive
+
+      - name: Install uv and Python
+        uses: astral-sh/setup-uv@v6
+        with:
+          python-version: '3.11'
+
+      - name: Run pre-commit with only lint group (no project deps)
+        run: |
+          uv run --only-group lint pre-commit run --all-files --show-diff-on-failure
+
+
+  build:
+    needs: lint
+    name: Build ${{ matrix.os }} Python ${{ matrix.python }}
+    strategy:
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            python: '3.9'
+          - os: ubuntu-22.04
+            python: '3.10'
+          - os: ubuntu-22.04
+            python: '3.11'
+          - os: ubuntu-22.04
+            python: '3.12'
+          - os: ubuntu-22.04
+            python: '3.13'
+          # ARM64 Linux builds
+          - os: ubuntu-24.04-arm
+            python: '3.9'
+          - os: ubuntu-24.04-arm
+            python: '3.10'
+          - os: ubuntu-24.04-arm
+            python: '3.11'
+          - os: ubuntu-24.04-arm
+            python: '3.12'
+          - os: ubuntu-24.04-arm
+            python: '3.13'
+          - os: macos-14
+            python: '3.9'
+          - os: macos-14
+            python: '3.10'
+          - os: macos-14
+            python: '3.11'
+          - os: macos-14
+            python: '3.12'
+          - os: macos-14
+            python: '3.13'
+          - os: macos-15
+            python: '3.9'
+          - os: macos-15
+            python: '3.10'
+          - os: macos-15
+            python: '3.11'
+          - os: macos-15
+            python: '3.12'
+          - os: macos-15
+            python: '3.13'
+          - os: macos-13
+            python: '3.9'
+          - os: macos-13
+            python: '3.10'
+          - os: macos-13
+            python: '3.11'
+          - os: macos-13
+            python: '3.12'
+          # Note: macos-13 + Python 3.13 excluded due to PyTorch compatibility
+          # (PyTorch 2.5+ supports Python 3.13 but not Intel Mac x86_64)
+    runs-on: ${{ matrix.os }}
+
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          ref: ${{ inputs.ref }}
+          submodules: recursive
+
+      - name: Install uv and Python
+        uses: astral-sh/setup-uv@v6
+        with:
+          python-version: ${{ matrix.python }}
+
+      - name: Install system dependencies (Ubuntu)
+        if: runner.os == 'Linux'
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y libomp-dev libboost-all-dev protobuf-compiler libzmq3-dev \
+            pkg-config libabsl-dev libaio-dev libprotobuf-dev \
+            patchelf
+
+          # Debug: Show system information
+          echo "🔍 System Information:"
+          echo "Architecture: $(uname -m)"
+          echo "OS: $(uname -a)"
+          echo "CPU info: $(lscpu | head -5)"
+
+          # Install math library based on architecture
+          ARCH=$(uname -m)
+          echo "🔍 Setting up math library for architecture: $ARCH"
+
+          if [[ "$ARCH" == "x86_64" ]]; then
+            # Install Intel MKL for DiskANN on x86_64
+            echo "📦 Installing Intel MKL for x86_64..."
+            wget -q https://registrationcenter-download.intel.com/akdlm/IRC_NAS/79153e0f-74d7-45af-b8c2-258941adf58a/intel-onemkl-2025.0.0.940.sh
+            sudo sh intel-onemkl-2025.0.0.940.sh -a --components intel.oneapi.lin.mkl.devel --action install --eula accept -s
+            source /opt/intel/oneapi/setvars.sh
+            echo "MKLROOT=/opt/intel/oneapi/mkl/latest" >> $GITHUB_ENV
+            echo "LD_LIBRARY_PATH=/opt/intel/oneapi/compiler/latest/linux/compiler/lib/intel64_lin" >> $GITHUB_ENV
+            echo "LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/intel/oneapi/mkl/latest/lib/intel64" >> $GITHUB_ENV
+            echo "✅ Intel MKL installed for x86_64"
+
+            # Debug: Check MKL installation
+            echo "🔍 MKL Installation Check:"
+            ls -la /opt/intel/oneapi/mkl/latest/ || echo "MKL directory not found"
+            ls -la /opt/intel/oneapi/mkl/latest/lib/ || echo "MKL lib directory not found"
+
+          elif [[ "$ARCH" == "aarch64" ]]; then
+            # Use OpenBLAS for ARM64 (MKL installer not compatible with ARM64)
+            echo "📦 Installing OpenBLAS for ARM64..."
+            sudo apt-get install -y libopenblas-dev liblapack-dev liblapacke-dev
+            echo "✅ OpenBLAS installed for ARM64"
+
+            # Debug: Check OpenBLAS installation
+            echo "🔍 OpenBLAS Installation Check:"
+            dpkg -l | grep openblas || echo "OpenBLAS package not found"
+            ls -la /usr/lib/aarch64-linux-gnu/openblas/ || echo "OpenBLAS directory not found"
+          fi
+
+          # Debug: Show final library paths
+          echo "🔍 Final LD_LIBRARY_PATH: $LD_LIBRARY_PATH"
+
+      - name: Install system dependencies (macOS)
+        if: runner.os == 'macOS'
+        run: |
+          # Don't install LLVM, use system clang for better compatibility
+          brew install libomp boost protobuf zeromq
+
+      - name: Install build dependencies
+        run: |
+          uv python install ${{ matrix.python }}
+          uv venv --python ${{ matrix.python }} .uv-build
+          if [[ "$RUNNER_OS" == "Windows" ]]; then
+            BUILD_PY=".uv-build\\Scripts\\python.exe"
+          else
+            BUILD_PY=".uv-build/bin/python"
+          fi
+          uv pip install --python "$BUILD_PY" scikit-build-core numpy swig Cython pybind11
+          if [[ "$RUNNER_OS" == "Linux" ]]; then
+            uv pip install --python "$BUILD_PY" auditwheel
+          else
+            uv pip install --python "$BUILD_PY" delocate
+          fi
+
+          if [[ "$RUNNER_OS" == "Windows" ]]; then
+            echo "$(pwd)\\.uv-build\\Scripts" >> $GITHUB_PATH
+          else
+            echo "$(pwd)/.uv-build/bin" >> $GITHUB_PATH
+          fi
+
+      - name: Set macOS environment variables
+        if: runner.os == 'macOS'
+        run: |
+          # Use brew --prefix to automatically detect Homebrew installation path
+          HOMEBREW_PREFIX=$(brew --prefix)
+          echo "HOMEBREW_PREFIX=${HOMEBREW_PREFIX}" >> $GITHUB_ENV
+          echo "OpenMP_ROOT=${HOMEBREW_PREFIX}/opt/libomp" >> $GITHUB_ENV
+
+          # Set CMAKE_PREFIX_PATH to let CMake find all packages automatically
+          echo "CMAKE_PREFIX_PATH=${HOMEBREW_PREFIX}" >> $GITHUB_ENV
+
+          # Set compiler flags for OpenMP (required for both backends)
+          echo "LDFLAGS=-L${HOMEBREW_PREFIX}/opt/libomp/lib" >> $GITHUB_ENV
+          echo "CPPFLAGS=-I${HOMEBREW_PREFIX}/opt/libomp/include" >> $GITHUB_ENV
+
+      - name: Build packages
+        run: |
+          # Build core (platform independent)
+          cd packages/leann-core
+          uv build
+          cd ../..
+
+          # Build HNSW backend
+          cd packages/leann-backend-hnsw
+          if [[ "${{ matrix.os }}" == macos-* ]]; then
+            # Use system clang for better compatibility
+            export CC=clang
+            export CXX=clang++
+            # Homebrew libraries on each macOS version require matching minimum version
+            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=13.0
+            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=14.0
+            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=15.0
+            fi
+            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
+          else
+            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
+          fi
+          cd ../..
+
+          # Build DiskANN backend
+          cd packages/leann-backend-diskann
+          if [[ "${{ matrix.os }}" == macos-* ]]; then
+            # Use system clang for better compatibility
+            export CC=clang
+            export CXX=clang++
+            # DiskANN requires macOS 13.3+ for sgesdd_ LAPACK function
+            # But Homebrew libraries on each macOS version require matching minimum version
+            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=13.3
+            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=14.0
+            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=15.0
+            fi
+            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
+          else
+            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
+          fi
+          cd ../..
+
+          # Build meta package (platform independent)
+          cd packages/leann
+          uv build
+          cd ../..
+
+      - name: Repair wheels (Linux)
+        if: runner.os == 'Linux'
+        run: |
+          # Repair HNSW wheel
+          cd packages/leann-backend-hnsw
+          if [ -d dist ]; then
+            auditwheel repair dist/*.whl -w dist_repaired
+            rm -rf dist
+            mv dist_repaired dist
+          fi
+          cd ../..
+
+          # Repair DiskANN wheel
+          cd packages/leann-backend-diskann
+          if [ -d dist ]; then
+            auditwheel repair dist/*.whl -w dist_repaired
+            rm -rf dist
+            mv dist_repaired dist
+          fi
+          cd ../..
+
+      - name: Repair wheels (macOS)
+        if: runner.os == 'macOS'
+        run: |
+          # Determine deployment target based on runner OS
+          # Must match the Homebrew libraries for each macOS version
+          if [[ "${{ matrix.os }}" == "macos-13" ]]; then
+            HNSW_TARGET="13.0"
+            DISKANN_TARGET="13.3"
+          elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
+            HNSW_TARGET="14.0"
+            DISKANN_TARGET="14.0"
+          elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+            HNSW_TARGET="15.0"
+            DISKANN_TARGET="15.0"
+          fi
+
+          # Repair HNSW wheel
+          cd packages/leann-backend-hnsw
+          if [ -d dist ]; then
+            export MACOSX_DEPLOYMENT_TARGET=$HNSW_TARGET
+            delocate-wheel -w dist_repaired -v --require-target-macos-version $HNSW_TARGET dist/*.whl
+            rm -rf dist
+            mv dist_repaired dist
+          fi
+          cd ../..
+
+          # Repair DiskANN wheel
+          cd packages/leann-backend-diskann
+          if [ -d dist ]; then
+            export MACOSX_DEPLOYMENT_TARGET=$DISKANN_TARGET
+            delocate-wheel -w dist_repaired -v --require-target-macos-version $DISKANN_TARGET dist/*.whl
+            rm -rf dist
+            mv dist_repaired dist
+          fi
+          cd ../..
+
+      - name: List built packages
+        run: |
+          echo "📦 Built packages:"
+          find packages/*/dist -name "*.whl" -o -name "*.tar.gz" | sort
+
+
+      - name: Install built packages for testing
+        run: |
+          # Create uv-managed virtual environment with the requested interpreter
+          uv python install ${{ matrix.python }}
+          uv venv --python ${{ matrix.python }}
+          source .venv/bin/activate || source .venv/Scripts/activate
+
+          if [[ "$RUNNER_OS" == "Windows" ]]; then
+            UV_PY=".venv\\Scripts\\python.exe"
+          else
+            UV_PY=".venv/bin/python"
+          fi
+
+          # Install test dependency group only (avoids reinstalling project package)
+          uv pip install --python "$UV_PY" --group test
+
+          # Install core wheel built in this job
+          CORE_WHL=$(find packages/leann-core/dist -maxdepth 1 -name "*.whl" -print -quit)
+          if [[ -n "$CORE_WHL" ]]; then
+            uv pip install --python "$UV_PY" "$CORE_WHL"
+          else
+            uv pip install --python "$UV_PY" packages/leann-core/dist/*.tar.gz
+          fi
+
+          PY_TAG=$($UV_PY -c "import sys; print(f'cp{sys.version_info[0]}{sys.version_info[1]}')")
+
+          if [[ "$RUNNER_OS" == "macOS" ]]; then
+            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=13.3
+            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=14.0
+            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=15.0
+            fi
+          fi
+
+          HNSW_WHL=$(find packages/leann-backend-hnsw/dist -maxdepth 1 -name "*-${PY_TAG}-*.whl" -print -quit)
+          if [[ -z "$HNSW_WHL" ]]; then
+            HNSW_WHL=$(find packages/leann-backend-hnsw/dist -maxdepth 1 -name "*-py3-*.whl" -print -quit)
+          fi
+          if [[ -n "$HNSW_WHL" ]]; then
+            uv pip install --python "$UV_PY" "$HNSW_WHL"
+          else
+            uv pip install --python "$UV_PY" ./packages/leann-backend-hnsw
+          fi
+
+          DISKANN_WHL=$(find packages/leann-backend-diskann/dist -maxdepth 1 -name "*-${PY_TAG}-*.whl" -print -quit)
+          if [[ -z "$DISKANN_WHL" ]]; then
+            DISKANN_WHL=$(find packages/leann-backend-diskann/dist -maxdepth 1 -name "*-py3-*.whl" -print -quit)
+          fi
+          if [[ -n "$DISKANN_WHL" ]]; then
+            uv pip install --python "$UV_PY" "$DISKANN_WHL"
+          else
+            uv pip install --python "$UV_PY" ./packages/leann-backend-diskann
+          fi
+
+          LEANN_WHL=$(find packages/leann/dist -maxdepth 1 -name "*.whl" -print -quit)
+          if [[ -n "$LEANN_WHL" ]]; then
+            uv pip install --python "$UV_PY" "$LEANN_WHL"
+          else
+            uv pip install --python "$UV_PY" packages/leann/dist/*.tar.gz
+          fi
+
+      - name: Run tests with pytest
+        env:
+          CI: true
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          HF_HUB_DISABLE_SYMLINKS: 1
+          TOKENIZERS_PARALLELISM: false
+          PYTORCH_ENABLE_MPS_FALLBACK: 0
+          OMP_NUM_THREADS: 1
+          MKL_NUM_THREADS: 1
+        run: |
+          source .venv/bin/activate || source .venv/Scripts/activate
+          pytest tests/ -v --tb=short
+
+      - name: Run sanity checks (optional)
+        run: |
+          # Activate virtual environment
+          source .venv/bin/activate || source .venv/Scripts/activate
+
+          # Run distance function tests if available
+          if [ -f test/sanity_checks/test_distance_functions.py ]; then
+            echo "Running distance function sanity checks..."
+            python test/sanity_checks/test_distance_functions.py || echo "⚠️ Distance function test failed, continuing..."
+          fi
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: packages-${{ matrix.os }}-py${{ matrix.python }}
+          path: packages/*/dist/
+
+
+  arch-smoke:
+    name: Arch Linux smoke test (install & import)
+    needs: build
+    runs-on: ubuntu-latest
+    container:
+      image: archlinux:latest
+
+    steps:
+      - name: Prepare system
+        run: |
+          pacman -Syu --noconfirm
+          pacman -S --noconfirm python python-pip gcc git zlib openssl
+
+      - name: Download ALL wheel artifacts from this run
+        uses: actions/download-artifact@v5
+        with:
+          # Don't specify name, download all artifacts
+          path: ./wheels
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v6
+
+      - name: Create virtual environment and install wheels
+        run: |
+          uv venv
+          source .venv/bin/activate || source .venv/Scripts/activate
+          uv pip install --find-links wheels leann-core
+          uv pip install --find-links wheels leann-backend-hnsw
+          uv pip install --find-links wheels leann-backend-diskann
+          uv pip install --find-links wheels leann
+
+      - name: Import & tiny runtime check
+        env:
+          OMP_NUM_THREADS: 1
+          MKL_NUM_THREADS: 1
+        run: |
+          source .venv/bin/activate || source .venv/Scripts/activate
+          python - <<'PY'
+          import leann
+          import leann_backend_hnsw as h
+          import leann_backend_diskann as d
+          from leann import LeannBuilder, LeannSearcher
+          b = LeannBuilder(backend_name="hnsw")
+          b.add_text("hello arch")
+          b.build_index("arch_demo.leann")
+          s = LeannSearcher("arch_demo.leann")
+          print("search:", s.search("hello", top_k=1))
+          PY

.github/workflows/link-check.yml

@@ -0,0 +1,19 @@
+name: Link Check
+
+on:
+  push:
+    branches: [ main, master ]
+  pull_request:
+  schedule:
+    - cron: "0 3 * * 1"
+
+jobs:
+  link-check:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: lycheeverse/lychee-action@v2
+        with:
+          args: --no-progress --insecure --user-agent 'curl/7.68.0' README.md docs/ apps/ examples/ benchmarks/
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/release-manual.yml

@@ -0,0 +1,129 @@
+name: Release
+
+on:
+  workflow_dispatch:
+    inputs:
+      version:
+        description: 'Version to release (e.g., 0.1.2)'
+        required: true
+        type: string
+
+jobs:
+  update-version:
+    name: Update Version
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    outputs:
+      commit-sha: ${{ steps.push.outputs.commit-sha }}
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Validate version
+        run: |
+          # Remove 'v' prefix if present for validation
+          VERSION_CLEAN="${{ inputs.version }}"
+          VERSION_CLEAN="${VERSION_CLEAN#v}"
+          if ! [[ "$VERSION_CLEAN" =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
+            echo "❌ Invalid version format. Expected format: X.Y.Z or vX.Y.Z"
+            exit 1
+          fi
+          echo "✅ Version format valid: ${{ inputs.version }}"
+
+      - name: Update versions and push
+        id: push
+        run: |
+          # Check current version
+          CURRENT_VERSION=$(grep "^version" packages/leann-core/pyproject.toml | cut -d'"' -f2)
+          echo "Current version: $CURRENT_VERSION"
+          echo "Target version: ${{ inputs.version }}"
+
+          if [ "$CURRENT_VERSION" = "${{ inputs.version }}" ]; then
+            echo "⚠️  Version is already ${{ inputs.version }}, skipping update"
+            COMMIT_SHA=$(git rev-parse HEAD)
+          else
+            ./scripts/bump_version.sh ${{ inputs.version }}
+            git config user.name "GitHub Actions"
+            git config user.email "actions@github.com"
+            git add packages/*/pyproject.toml
+            git commit -m "chore: release v${{ inputs.version }}"
+            git push origin main
+            COMMIT_SHA=$(git rev-parse HEAD)
+            echo "✅ Pushed version update: $COMMIT_SHA"
+          fi
+
+          echo "commit-sha=$COMMIT_SHA" >> $GITHUB_OUTPUT
+
+  build-packages:
+    name: Build packages
+    needs: update-version
+    uses: ./.github/workflows/build-reusable.yml
+    with:
+      ref: 'main'
+
+  publish:
+    name: Publish and Release
+    needs: [update-version, build-packages]
+    if: always() && needs.update-version.result == 'success' && needs.build-packages.result == 'success'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: 'main'
+
+      - name: Download all artifacts
+        uses: actions/download-artifact@v4
+        with:
+          path: dist-artifacts
+
+      - name: Collect packages
+        run: |
+          mkdir -p dist
+          find dist-artifacts -name "*.whl" -exec cp {} dist/ \;
+          find dist-artifacts -name "*.tar.gz" -exec cp {} dist/ \;
+
+          echo "📦 Packages to publish:"
+          ls -la dist/
+
+      - name: Publish to PyPI
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
+        run: |
+          if [ -z "$TWINE_PASSWORD" ]; then
+            echo "❌ PYPI_API_TOKEN not configured!"
+            exit 1
+          fi
+
+          pip install twine
+          twine upload dist/* --skip-existing --verbose
+
+          echo "✅ Published to PyPI!"
+
+      - name: Create release
+        run: |
+          # Check if tag already exists
+          if git rev-parse "v${{ inputs.version }}" >/dev/null 2>&1; then
+            echo "⚠️  Tag v${{ inputs.version }} already exists, skipping tag creation"
+          else
+            git tag "v${{ inputs.version }}"
+            git push origin "v${{ inputs.version }}"
+            echo "✅ Created and pushed tag v${{ inputs.version }}"
+          fi
+
+          # Check if release already exists
+          if gh release view "v${{ inputs.version }}" >/dev/null 2>&1; then
+            echo "⚠️  Release v${{ inputs.version }} already exists, skipping release creation"
+          else
+            gh release create "v${{ inputs.version }}" \
+              --title "Release v${{ inputs.version }}" \
+              --notes "🚀 Released to PyPI: https://pypi.org/project/leann/${{ inputs.version }}/" \
+              --latest
+            echo "✅ Created GitHub release v${{ inputs.version }}"
+          fi
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.gitignore

@@ -9,19 +9,21 @@ demo/indices/
 outputs/
 *.pkl
 *.pdf
-*.idx 
+*.idx
 *.map
 .history/
-scripts/
 lm_eval.egg-info/
 demo/experiment_results/**/*.json
 *.jsonl
 *.eml
 *.emlx
 *.json
+*.png
+!.vscode/*.json
 *.sh
 *.txt
 !CMakeLists.txt
+!llms.txt
 latency_breakdown*.json
 experiment_results/eval_results/diskann/*.json
 aws/
@@ -35,11 +37,15 @@ build/
 nprobe_logs/
 micro/results
 micro/contriever-INT8
-examples/data/*
-!examples/data/2501.14312v1 (1).pdf
-!examples/data/2506.08276v1.pdf
-!examples/data/PrideandPrejudice.txt
-!examples/data/README.md
+data/*
+!data/2501.14312v1 (1).pdf
+!data/2506.08276v1.pdf
+!data/PrideandPrejudice.txt
+!data/huawei_pangu.md
+!data/ground_truth/
+!data/indices/
+!data/queries/
+!data/.gitattributes
 *.qdstrm
 benchmark_results/
 results/
@@ -86,4 +92,16 @@ packages/leann-backend-diskann/third_party/DiskANN/_deps/
 *.meta.json
 *.passages.json
 
-batchtest.py
+batchtest.py
+tests/__pytest_cache__/
+tests/__pycache__/
+benchmarks/data/
+
+## multi vector
+apps/multimodal/vision-based-pdf-multi-vector/multi-vector-colpali-native-weaviate.py
+
+# Ignore all PDFs (keep data exceptions above) and do not track demo PDFs
+# If you need to commit a specific demo PDF, remove this negation locally.
+# The following line used to force-add a large demo PDF; remove it to satisfy pre-commit:
+# !apps/multimodal/vision-based-pdf-multi-vector/pdfs/2004.12832v2.pdf
+!apps/multimodal/vision-based-pdf-multi-vector/fig/*

.gitmodules

@@ -14,3 +14,6 @@
 [submodule "packages/leann-backend-hnsw/third_party/libzmq"]
 	path = packages/leann-backend-hnsw/third_party/libzmq
 	url = https://github.com/zeromq/libzmq.git
+[submodule "packages/astchunk-leann"]
+	path = packages/astchunk-leann
+	url = https://github.com/yichuan-w/astchunk-leann.git

.pre-commit-config.yaml

@@ -0,0 +1,17 @@
+repos:
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v5.0.0
+    hooks:
+      - id: trailing-whitespace
+      - id: end-of-file-fixer
+      - id: check-yaml
+      - id: check-added-large-files
+      - id: check-merge-conflict
+      - id: debug-statements
+
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.12.7  # Fixed version to match pyproject.toml
+    hooks:
+      - id: ruff
+        args: [--fix, --exit-non-zero-on-fix]
+      - id: ruff-format

.vscode/extensions.json

@@ -0,0 +1,5 @@
+{
+    "recommendations": [
+        "charliermarsh.ruff",
+    ]
+}

.vscode/settings.json

@@ -0,0 +1,22 @@
+{
+  "python.defaultInterpreterPath": ".venv/bin/python",
+  "python.terminal.activateEnvironment": true,
+  "[python]": {
+    "editor.defaultFormatter": "charliermarsh.ruff",
+    "editor.formatOnSave": true,
+    "editor.codeActionsOnSave": {
+      "source.organizeImports": "explicit",
+      "source.fixAll": "explicit"
+    },
+    "editor.insertSpaces": true,
+    "editor.tabSize": 4
+  },
+  "ruff.enable": true,
+  "files.watcherExclude": {
+    "**/.venv/**": true,
+    "**/__pycache__/**": true,
+    "**/*.egg-info/**": true,
+    "**/build/**": true,
+    "**/dist/**": true
+  }
+}

apps/base_rag_example.py

@@ -0,0 +1,377 @@
+"""
+Base class for unified RAG examples interface.
+Provides common parameters and functionality for all RAG examples.
+"""
+
+import argparse
+from abc import ABC, abstractmethod
+from pathlib import Path
+from typing import Any
+
+import dotenv
+from leann.api import LeannBuilder, LeannChat
+from leann.interactive_utils import create_rag_session
+from leann.registry import register_project_directory
+from leann.settings import resolve_ollama_host, resolve_openai_api_key, resolve_openai_base_url
+
+dotenv.load_dotenv()
+
+
+class BaseRAGExample(ABC):
+    """Base class for all RAG examples with unified interface."""
+
+    def __init__(
+        self,
+        name: str,
+        description: str,
+        default_index_name: str,
+    ):
+        self.name = name
+        self.description = description
+        self.default_index_name = default_index_name
+        self.parser = self._create_parser()
+
+    def _create_parser(self) -> argparse.ArgumentParser:
+        """Create argument parser with common parameters."""
+        parser = argparse.ArgumentParser(
+            description=self.description, formatter_class=argparse.RawDescriptionHelpFormatter
+        )
+
+        # Core parameters (all examples share these)
+        core_group = parser.add_argument_group("Core Parameters")
+        core_group.add_argument(
+            "--index-dir",
+            type=str,
+            default=f"./{self.default_index_name}",
+            help=f"Directory to store the index (default: ./{self.default_index_name})",
+        )
+        core_group.add_argument(
+            "--query",
+            type=str,
+            default=None,
+            help="Query to run (if not provided, will run in interactive mode)",
+        )
+        # Allow subclasses to override default max_items
+        max_items_default = getattr(self, "max_items_default", -1)
+        core_group.add_argument(
+            "--max-items",
+            type=int,
+            default=max_items_default,
+            help="Maximum number of items to process  -1 for all, means index all documents, and you should set it to a reasonable number if you have a large dataset and try at the first time)",
+        )
+        core_group.add_argument(
+            "--force-rebuild", action="store_true", help="Force rebuild index even if it exists"
+        )
+
+        # Embedding parameters
+        embedding_group = parser.add_argument_group("Embedding Parameters")
+        # Allow subclasses to override default embedding_model
+        embedding_model_default = getattr(self, "embedding_model_default", "facebook/contriever")
+        embedding_group.add_argument(
+            "--embedding-model",
+            type=str,
+            default=embedding_model_default,
+            help=f"Embedding model to use (default: {embedding_model_default}), we provide facebook/contriever, text-embedding-3-small,mlx-community/Qwen3-Embedding-0.6B-8bit or nomic-embed-text",
+        )
+        embedding_group.add_argument(
+            "--embedding-mode",
+            type=str,
+            default="sentence-transformers",
+            choices=["sentence-transformers", "openai", "mlx", "ollama"],
+            help="Embedding backend mode (default: sentence-transformers), we provide sentence-transformers, openai, mlx, or ollama",
+        )
+        embedding_group.add_argument(
+            "--embedding-host",
+            type=str,
+            default=None,
+            help="Override Ollama-compatible embedding host",
+        )
+        embedding_group.add_argument(
+            "--embedding-api-base",
+            type=str,
+            default=None,
+            help="Base URL for OpenAI-compatible embedding services",
+        )
+        embedding_group.add_argument(
+            "--embedding-api-key",
+            type=str,
+            default=None,
+            help="API key for embedding service (defaults to OPENAI_API_KEY)",
+        )
+
+        # LLM parameters
+        llm_group = parser.add_argument_group("LLM Parameters")
+        llm_group.add_argument(
+            "--llm",
+            type=str,
+            default="openai",
+            choices=["openai", "ollama", "hf", "simulated"],
+            help="LLM backend: openai, ollama, or hf (default: openai)",
+        )
+        llm_group.add_argument(
+            "--llm-model",
+            type=str,
+            default=None,
+            help="Model name (default: gpt-4o) e.g., gpt-4o-mini, llama3.2:1b, Qwen/Qwen2.5-1.5B-Instruct",
+        )
+        llm_group.add_argument(
+            "--llm-host",
+            type=str,
+            default=None,
+            help="Host for Ollama-compatible APIs (defaults to LEANN_OLLAMA_HOST/OLLAMA_HOST)",
+        )
+        llm_group.add_argument(
+            "--thinking-budget",
+            type=str,
+            choices=["low", "medium", "high"],
+            default=None,
+            help="Thinking budget for reasoning models (low/medium/high). Supported by GPT-Oss:20b and other reasoning models.",
+        )
+        llm_group.add_argument(
+            "--llm-api-base",
+            type=str,
+            default=None,
+            help="Base URL for OpenAI-compatible APIs",
+        )
+        llm_group.add_argument(
+            "--llm-api-key",
+            type=str,
+            default=None,
+            help="API key for OpenAI-compatible APIs (defaults to OPENAI_API_KEY)",
+        )
+
+        # AST Chunking parameters
+        ast_group = parser.add_argument_group("AST Chunking Parameters")
+        ast_group.add_argument(
+            "--use-ast-chunking",
+            action="store_true",
+            help="Enable AST-aware chunking for code files (requires astchunk)",
+        )
+        ast_group.add_argument(
+            "--ast-chunk-size",
+            type=int,
+            default=512,
+            help="Maximum characters per AST chunk (default: 512)",
+        )
+        ast_group.add_argument(
+            "--ast-chunk-overlap",
+            type=int,
+            default=64,
+            help="Overlap between AST chunks (default: 64)",
+        )
+        ast_group.add_argument(
+            "--code-file-extensions",
+            nargs="+",
+            default=None,
+            help="Additional code file extensions to process with AST chunking (e.g., .py .java .cs .ts)",
+        )
+        ast_group.add_argument(
+            "--ast-fallback-traditional",
+            action="store_true",
+            default=True,
+            help="Fall back to traditional chunking if AST chunking fails (default: True)",
+        )
+
+        # Search parameters
+        search_group = parser.add_argument_group("Search Parameters")
+        search_group.add_argument(
+            "--top-k", type=int, default=20, help="Number of results to retrieve (default: 20)"
+        )
+        search_group.add_argument(
+            "--search-complexity",
+            type=int,
+            default=32,
+            help="Search complexity for graph traversal (default: 64)",
+        )
+
+        # Index building parameters
+        index_group = parser.add_argument_group("Index Building Parameters")
+        index_group.add_argument(
+            "--backend-name",
+            type=str,
+            default="hnsw",
+            choices=["hnsw", "diskann"],
+            help="Backend to use for index (default: hnsw)",
+        )
+        index_group.add_argument(
+            "--graph-degree",
+            type=int,
+            default=32,
+            help="Graph degree for index construction (default: 32)",
+        )
+        index_group.add_argument(
+            "--build-complexity",
+            type=int,
+            default=64,
+            help="Build complexity for index construction (default: 64)",
+        )
+        index_group.add_argument(
+            "--no-compact",
+            action="store_true",
+            help="Disable compact index storage",
+        )
+        index_group.add_argument(
+            "--no-recompute",
+            action="store_true",
+            help="Disable embedding recomputation",
+        )
+
+        # Add source-specific parameters
+        self._add_specific_arguments(parser)
+
+        return parser
+
+    @abstractmethod
+    def _add_specific_arguments(self, parser: argparse.ArgumentParser):
+        """Add source-specific arguments. Override in subclasses."""
+        pass
+
+    @abstractmethod
+    async def load_data(self, args) -> list[str]:
+        """Load data from the source. Returns list of text chunks."""
+        pass
+
+    def get_llm_config(self, args) -> dict[str, Any]:
+        """Get LLM configuration based on arguments."""
+        config = {"type": args.llm}
+
+        if args.llm == "openai":
+            config["model"] = args.llm_model or "gpt-4o"
+            config["base_url"] = resolve_openai_base_url(args.llm_api_base)
+            resolved_key = resolve_openai_api_key(args.llm_api_key)
+            if resolved_key:
+                config["api_key"] = resolved_key
+        elif args.llm == "ollama":
+            config["model"] = args.llm_model or "llama3.2:1b"
+            config["host"] = resolve_ollama_host(args.llm_host)
+        elif args.llm == "hf":
+            config["model"] = args.llm_model or "Qwen/Qwen2.5-1.5B-Instruct"
+        elif args.llm == "simulated":
+            # Simulated LLM doesn't need additional configuration
+            pass
+
+        return config
+
+    async def build_index(self, args, texts: list[str]) -> str:
+        """Build LEANN index from texts."""
+        index_path = str(Path(args.index_dir) / f"{self.default_index_name}.leann")
+
+        print(f"\n[Building Index] Creating {self.name} index...")
+        print(f"Total text chunks: {len(texts)}")
+
+        embedding_options: dict[str, Any] = {}
+        if args.embedding_mode == "ollama":
+            embedding_options["host"] = resolve_ollama_host(args.embedding_host)
+        elif args.embedding_mode == "openai":
+            embedding_options["base_url"] = resolve_openai_base_url(args.embedding_api_base)
+            resolved_embedding_key = resolve_openai_api_key(args.embedding_api_key)
+            if resolved_embedding_key:
+                embedding_options["api_key"] = resolved_embedding_key
+
+        builder = LeannBuilder(
+            backend_name=args.backend_name,
+            embedding_model=args.embedding_model,
+            embedding_mode=args.embedding_mode,
+            embedding_options=embedding_options or None,
+            graph_degree=args.graph_degree,
+            complexity=args.build_complexity,
+            is_compact=not args.no_compact,
+            is_recompute=not args.no_recompute,
+            num_threads=1,  # Force single-threaded mode
+        )
+
+        # Add texts in batches for better progress tracking
+        batch_size = 1000
+        for i in range(0, len(texts), batch_size):
+            batch = texts[i : i + batch_size]
+            for text in batch:
+                builder.add_text(text)
+            print(f"Added {min(i + batch_size, len(texts))}/{len(texts)} texts...")
+
+        print("Building index structure...")
+        builder.build_index(index_path)
+        print(f"Index saved to: {index_path}")
+
+        # Register project directory so leann list can discover this index
+        # The index is saved as args.index_dir/index_name.leann
+        # We want to register the current working directory where the app is run
+        register_project_directory(Path.cwd())
+
+        return index_path
+
+    async def run_interactive_chat(self, args, index_path: str):
+        """Run interactive chat with the index."""
+        chat = LeannChat(
+            index_path,
+            llm_config=self.get_llm_config(args),
+            system_prompt=f"You are a helpful assistant that answers questions about {self.name} data.",
+            complexity=args.search_complexity,
+        )
+
+        # Create interactive session
+        session = create_rag_session(
+            app_name=self.name.lower().replace(" ", "_"), data_description=self.name
+        )
+
+        def handle_query(query: str):
+            # Prepare LLM kwargs with thinking budget if specified
+            llm_kwargs = {}
+            if hasattr(args, "thinking_budget") and args.thinking_budget:
+                llm_kwargs["thinking_budget"] = args.thinking_budget
+
+            response = chat.ask(
+                query,
+                top_k=args.top_k,
+                complexity=args.search_complexity,
+                llm_kwargs=llm_kwargs,
+            )
+            print(f"\nAssistant: {response}\n")
+
+        session.run_interactive_loop(handle_query)
+
+    async def run_single_query(self, args, index_path: str, query: str):
+        """Run a single query against the index."""
+        chat = LeannChat(
+            index_path,
+            llm_config=self.get_llm_config(args),
+            complexity=args.search_complexity,
+        )
+
+        print(f"\n[Query]: \033[36m{query}\033[0m")
+
+        # Prepare LLM kwargs with thinking budget if specified
+        llm_kwargs = {}
+        if hasattr(args, "thinking_budget") and args.thinking_budget:
+            llm_kwargs["thinking_budget"] = args.thinking_budget
+
+        response = chat.ask(
+            query, top_k=args.top_k, complexity=args.search_complexity, llm_kwargs=llm_kwargs
+        )
+        print(f"\n[Response]: \033[36m{response}\033[0m")
+
+    async def run(self):
+        """Main entry point for the example."""
+        args = self.parser.parse_args()
+
+        # Check if index exists
+        index_path = str(Path(args.index_dir) / f"{self.default_index_name}.leann")
+        index_exists = Path(args.index_dir).exists()
+
+        if not index_exists or args.force_rebuild:
+            # Load data and build index
+            print(f"\n{'Rebuilding' if index_exists else 'Building'} index...")
+            texts = await self.load_data(args)
+
+            if not texts:
+                print("No data found to index!")
+                return
+
+            index_path = await self.build_index(args, texts)
+        else:
+            print(f"\nUsing existing index in {args.index_dir}")
+
+        # Run query or interactive mode
+        if args.query:
+            await self.run_single_query(args, index_path, args.query)
+        else:
+            await self.run_interactive_chat(args, index_path)

apps/browser_rag.py

@@ -0,0 +1,171 @@
+"""
+Browser History RAG example using the unified interface.
+Supports Chrome browser history.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+from base_rag_example import BaseRAGExample
+from chunking import create_text_chunks
+
+from .history_data.history import ChromeHistoryReader
+
+
+class BrowserRAG(BaseRAGExample):
+    """RAG example for Chrome browser history."""
+
+    def __init__(self):
+        # Set default values BEFORE calling super().__init__
+        self.embedding_model_default = (
+            "sentence-transformers/all-MiniLM-L6-v2"  # Fast 384-dim model
+        )
+
+        super().__init__(
+            name="Browser History",
+            description="Process and query Chrome browser history with LEANN",
+            default_index_name="google_history_index",
+        )
+
+    def _add_specific_arguments(self, parser):
+        """Add browser-specific arguments."""
+        browser_group = parser.add_argument_group("Browser Parameters")
+        browser_group.add_argument(
+            "--chrome-profile",
+            type=str,
+            default=None,
+            help="Path to Chrome profile directory (auto-detected if not specified)",
+        )
+        browser_group.add_argument(
+            "--auto-find-profiles",
+            action="store_true",
+            default=True,
+            help="Automatically find all Chrome profiles (default: True)",
+        )
+        browser_group.add_argument(
+            "--chunk-size", type=int, default=256, help="Text chunk size (default: 256)"
+        )
+        browser_group.add_argument(
+            "--chunk-overlap", type=int, default=128, help="Text chunk overlap (default: 128)"
+        )
+
+    def _get_chrome_base_path(self) -> Path:
+        """Get the base Chrome profile path based on OS."""
+        if sys.platform == "darwin":
+            return Path.home() / "Library" / "Application Support" / "Google" / "Chrome"
+        elif sys.platform.startswith("linux"):
+            return Path.home() / ".config" / "google-chrome"
+        elif sys.platform == "win32":
+            return Path(os.environ["LOCALAPPDATA"]) / "Google" / "Chrome" / "User Data"
+        else:
+            raise ValueError(f"Unsupported platform: {sys.platform}")
+
+    def _find_chrome_profiles(self) -> list[Path]:
+        """Auto-detect all Chrome profiles."""
+        base_path = self._get_chrome_base_path()
+        if not base_path.exists():
+            return []
+
+        profiles = []
+
+        # Check Default profile
+        default_profile = base_path / "Default"
+        if default_profile.exists() and (default_profile / "History").exists():
+            profiles.append(default_profile)
+
+        # Check numbered profiles
+        for item in base_path.iterdir():
+            if item.is_dir() and item.name.startswith("Profile "):
+                if (item / "History").exists():
+                    profiles.append(item)
+
+        return profiles
+
+    async def load_data(self, args) -> list[str]:
+        """Load browser history and convert to text chunks."""
+        # Determine Chrome profiles
+        if args.chrome_profile and not args.auto_find_profiles:
+            profile_dirs = [Path(args.chrome_profile)]
+        else:
+            print("Auto-detecting Chrome profiles...")
+            profile_dirs = self._find_chrome_profiles()
+
+            # If specific profile given, filter to just that one
+            if args.chrome_profile:
+                profile_path = Path(args.chrome_profile)
+                profile_dirs = [p for p in profile_dirs if p == profile_path]
+
+        if not profile_dirs:
+            print("No Chrome profiles found!")
+            print("Please specify --chrome-profile manually")
+            return []
+
+        print(f"Found {len(profile_dirs)} Chrome profiles")
+
+        # Create reader
+        reader = ChromeHistoryReader()
+
+        # Process each profile
+        all_documents = []
+        total_processed = 0
+
+        for i, profile_dir in enumerate(profile_dirs):
+            print(f"\nProcessing profile {i + 1}/{len(profile_dirs)}: {profile_dir.name}")
+
+            try:
+                # Apply max_items limit per profile
+                max_per_profile = -1
+                if args.max_items > 0:
+                    remaining = args.max_items - total_processed
+                    if remaining <= 0:
+                        break
+                    max_per_profile = remaining
+
+                # Load history
+                documents = reader.load_data(
+                    chrome_profile_path=str(profile_dir),
+                    max_count=max_per_profile,
+                )
+
+                if documents:
+                    all_documents.extend(documents)
+                    total_processed += len(documents)
+                    print(f"Processed {len(documents)} history entries from this profile")
+
+            except Exception as e:
+                print(f"Error processing {profile_dir}: {e}")
+                continue
+
+        if not all_documents:
+            print("No browser history found to process!")
+            return []
+
+        print(f"\nTotal history entries processed: {len(all_documents)}")
+
+        # Convert to text chunks
+        all_texts = create_text_chunks(
+            all_documents, chunk_size=args.chunk_size, chunk_overlap=args.chunk_overlap
+        )
+
+        return all_texts
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Example queries for browser history RAG
+    print("\n🌐 Browser History RAG Example")
+    print("=" * 50)
+    print("\nExample queries you can try:")
+    print("- 'What websites did I visit about machine learning?'")
+    print("- 'Find my search history about programming'")
+    print("- 'What YouTube videos did I watch recently?'")
+    print("- 'Show me websites about travel planning'")
+    print("\nNote: Make sure Chrome is closed before running\n")
+
+    rag = BrowserRAG()
+    asyncio.run(rag.run())

apps/chatgpt_data/chatgpt_reader.py

@@ -0,0 +1,413 @@
+"""
+ChatGPT export data reader.
+
+Reads and processes ChatGPT export data from chat.html files.
+"""
+
+import re
+from pathlib import Path
+from typing import Any
+from zipfile import ZipFile
+
+from bs4 import BeautifulSoup
+from llama_index.core import Document
+from llama_index.core.readers.base import BaseReader
+
+
+class ChatGPTReader(BaseReader):
+    """
+    ChatGPT export data reader.
+
+    Reads ChatGPT conversation data from exported chat.html files or zip archives.
+    Processes conversations into structured documents with metadata.
+    """
+
+    def __init__(self, concatenate_conversations: bool = True) -> None:
+        """
+        Initialize.
+
+        Args:
+            concatenate_conversations: Whether to concatenate messages within conversations for better context
+        """
+        try:
+            from bs4 import BeautifulSoup  # noqa
+        except ImportError:
+            raise ImportError("`beautifulsoup4` package not found: `pip install beautifulsoup4`")
+
+        self.concatenate_conversations = concatenate_conversations
+
+    def _extract_html_from_zip(self, zip_path: Path) -> str | None:
+        """
+        Extract chat.html from ChatGPT export zip file.
+
+        Args:
+            zip_path: Path to the ChatGPT export zip file
+
+        Returns:
+            HTML content as string, or None if not found
+        """
+        try:
+            with ZipFile(zip_path, "r") as zip_file:
+                # Look for chat.html or conversations.html
+                html_files = [
+                    f
+                    for f in zip_file.namelist()
+                    if f.endswith(".html") and ("chat" in f.lower() or "conversation" in f.lower())
+                ]
+
+                if not html_files:
+                    print(f"No HTML chat file found in {zip_path}")
+                    return None
+
+                # Use the first HTML file found
+                html_file = html_files[0]
+                print(f"Found HTML file: {html_file}")
+
+                with zip_file.open(html_file) as f:
+                    return f.read().decode("utf-8", errors="ignore")
+
+        except Exception as e:
+            print(f"Error extracting HTML from zip {zip_path}: {e}")
+            return None
+
+    def _parse_chatgpt_html(self, html_content: str) -> list[dict]:
+        """
+        Parse ChatGPT HTML export to extract conversations.
+
+        Args:
+            html_content: HTML content from ChatGPT export
+
+        Returns:
+            List of conversation dictionaries
+        """
+        soup = BeautifulSoup(html_content, "html.parser")
+        conversations = []
+
+        # Try different possible structures for ChatGPT exports
+        # Structure 1: Look for conversation containers
+        conversation_containers = soup.find_all(
+            ["div", "section"], class_=re.compile(r"conversation|chat", re.I)
+        )
+
+        if not conversation_containers:
+            # Structure 2: Look for message containers directly
+            conversation_containers = [soup]  # Use the entire document as one conversation
+
+        for container in conversation_containers:
+            conversation = self._extract_conversation_from_container(container)
+            if conversation and conversation.get("messages"):
+                conversations.append(conversation)
+
+        # If no structured conversations found, try to extract all text as one conversation
+        if not conversations:
+            all_text = soup.get_text(separator="\n", strip=True)
+            if all_text:
+                conversations.append(
+                    {
+                        "title": "ChatGPT Conversation",
+                        "messages": [{"role": "mixed", "content": all_text, "timestamp": None}],
+                        "timestamp": None,
+                    }
+                )
+
+        return conversations
+
+    def _extract_conversation_from_container(self, container) -> dict | None:
+        """
+        Extract conversation data from a container element.
+
+        Args:
+            container: BeautifulSoup element containing conversation
+
+        Returns:
+            Dictionary with conversation data or None
+        """
+        messages = []
+
+        # Look for message elements with various possible structures
+        message_selectors = ['[class*="message"]', '[class*="chat"]', "[data-message]", "p", "div"]
+
+        for selector in message_selectors:
+            message_elements = container.select(selector)
+            if message_elements:
+                break
+        else:
+            message_elements = []
+
+        # If no structured messages found, treat the entire container as one message
+        if not message_elements:
+            text_content = container.get_text(separator="\n", strip=True)
+            if text_content:
+                messages.append({"role": "mixed", "content": text_content, "timestamp": None})
+        else:
+            for element in message_elements:
+                message = self._extract_message_from_element(element)
+                if message:
+                    messages.append(message)
+
+        if not messages:
+            return None
+
+        # Try to extract conversation title
+        title_element = container.find(["h1", "h2", "h3", "title"])
+        title = title_element.get_text(strip=True) if title_element else "ChatGPT Conversation"
+
+        # Try to extract timestamp from various possible locations
+        timestamp = self._extract_timestamp_from_container(container)
+
+        return {"title": title, "messages": messages, "timestamp": timestamp}
+
+    def _extract_message_from_element(self, element) -> dict | None:
+        """
+        Extract message data from an element.
+
+        Args:
+            element: BeautifulSoup element containing message
+
+        Returns:
+            Dictionary with message data or None
+        """
+        text_content = element.get_text(separator=" ", strip=True)
+
+        # Skip empty or very short messages
+        if not text_content or len(text_content.strip()) < 3:
+            return None
+
+        # Try to determine role (user/assistant) from class names or content
+        role = "mixed"  # Default role
+
+        class_names = " ".join(element.get("class", [])).lower()
+        if "user" in class_names or "human" in class_names:
+            role = "user"
+        elif "assistant" in class_names or "ai" in class_names or "gpt" in class_names:
+            role = "assistant"
+        elif text_content.lower().startswith(("you:", "user:", "me:")):
+            role = "user"
+            text_content = re.sub(r"^(you|user|me):\s*", "", text_content, flags=re.IGNORECASE)
+        elif text_content.lower().startswith(("chatgpt:", "assistant:", "ai:")):
+            role = "assistant"
+            text_content = re.sub(
+                r"^(chatgpt|assistant|ai):\s*", "", text_content, flags=re.IGNORECASE
+            )
+
+        # Try to extract timestamp
+        timestamp = self._extract_timestamp_from_element(element)
+
+        return {"role": role, "content": text_content, "timestamp": timestamp}
+
+    def _extract_timestamp_from_element(self, element) -> str | None:
+        """Extract timestamp from element."""
+        # Look for timestamp in various attributes and child elements
+        timestamp_attrs = ["data-timestamp", "timestamp", "datetime"]
+        for attr in timestamp_attrs:
+            if element.get(attr):
+                return element.get(attr)
+
+        # Look for time elements
+        time_element = element.find("time")
+        if time_element:
+            return time_element.get("datetime") or time_element.get_text(strip=True)
+
+        # Look for date-like text patterns
+        text = element.get_text()
+        date_patterns = [r"\d{4}-\d{2}-\d{2}", r"\d{1,2}/\d{1,2}/\d{4}", r"\w+ \d{1,2}, \d{4}"]
+
+        for pattern in date_patterns:
+            match = re.search(pattern, text)
+            if match:
+                return match.group()
+
+        return None
+
+    def _extract_timestamp_from_container(self, container) -> str | None:
+        """Extract timestamp from conversation container."""
+        return self._extract_timestamp_from_element(container)
+
+    def _create_concatenated_content(self, conversation: dict) -> str:
+        """
+        Create concatenated content from conversation messages.
+
+        Args:
+            conversation: Dictionary containing conversation data
+
+        Returns:
+            Formatted concatenated content
+        """
+        title = conversation.get("title", "ChatGPT Conversation")
+        messages = conversation.get("messages", [])
+        timestamp = conversation.get("timestamp", "Unknown")
+
+        # Build message content
+        message_parts = []
+        for message in messages:
+            role = message.get("role", "mixed")
+            content = message.get("content", "")
+            msg_timestamp = message.get("timestamp", "")
+
+            if role == "user":
+                prefix = "[You]"
+            elif role == "assistant":
+                prefix = "[ChatGPT]"
+            else:
+                prefix = "[Message]"
+
+            # Add timestamp if available
+            if msg_timestamp:
+                prefix += f" ({msg_timestamp})"
+
+            message_parts.append(f"{prefix}: {content}")
+
+        concatenated_text = "\n\n".join(message_parts)
+
+        # Create final document content
+        doc_content = f"""Conversation: {title}
+Date: {timestamp}
+Messages ({len(messages)} messages):
+
+{concatenated_text}
+"""
+        return doc_content
+
+    def load_data(self, input_dir: str | None = None, **load_kwargs: Any) -> list[Document]:
+        """
+        Load ChatGPT export data.
+
+        Args:
+            input_dir: Directory containing ChatGPT export files or path to specific file
+            **load_kwargs:
+                max_count (int): Maximum number of conversations to process
+                chatgpt_export_path (str): Specific path to ChatGPT export file/directory
+                include_metadata (bool): Whether to include metadata in documents
+        """
+        docs: list[Document] = []
+        max_count = load_kwargs.get("max_count", -1)
+        chatgpt_export_path = load_kwargs.get("chatgpt_export_path", input_dir)
+        include_metadata = load_kwargs.get("include_metadata", True)
+
+        if not chatgpt_export_path:
+            print("No ChatGPT export path provided")
+            return docs
+
+        export_path = Path(chatgpt_export_path)
+
+        if not export_path.exists():
+            print(f"ChatGPT export path not found: {export_path}")
+            return docs
+
+        html_content = None
+
+        # Handle different input types
+        if export_path.is_file():
+            if export_path.suffix.lower() == ".zip":
+                # Extract HTML from zip file
+                html_content = self._extract_html_from_zip(export_path)
+            elif export_path.suffix.lower() == ".html":
+                # Read HTML file directly
+                try:
+                    with open(export_path, encoding="utf-8", errors="ignore") as f:
+                        html_content = f.read()
+                except Exception as e:
+                    print(f"Error reading HTML file {export_path}: {e}")
+                    return docs
+            else:
+                print(f"Unsupported file type: {export_path.suffix}")
+                return docs
+
+        elif export_path.is_dir():
+            # Look for HTML files in directory
+            html_files = list(export_path.glob("*.html"))
+            zip_files = list(export_path.glob("*.zip"))
+
+            if html_files:
+                # Use first HTML file found
+                html_file = html_files[0]
+                print(f"Found HTML file: {html_file}")
+                try:
+                    with open(html_file, encoding="utf-8", errors="ignore") as f:
+                        html_content = f.read()
+                except Exception as e:
+                    print(f"Error reading HTML file {html_file}: {e}")
+                    return docs
+
+            elif zip_files:
+                # Use first zip file found
+                zip_file = zip_files[0]
+                print(f"Found zip file: {zip_file}")
+                html_content = self._extract_html_from_zip(zip_file)
+
+            else:
+                print(f"No HTML or zip files found in {export_path}")
+                return docs
+
+        if not html_content:
+            print("No HTML content found to process")
+            return docs
+
+        # Parse conversations from HTML
+        print("Parsing ChatGPT conversations from HTML...")
+        conversations = self._parse_chatgpt_html(html_content)
+
+        if not conversations:
+            print("No conversations found in HTML content")
+            return docs
+
+        print(f"Found {len(conversations)} conversations")
+
+        # Process conversations into documents
+        count = 0
+        for conversation in conversations:
+            if max_count > 0 and count >= max_count:
+                break
+
+            if self.concatenate_conversations:
+                # Create one document per conversation with concatenated messages
+                doc_content = self._create_concatenated_content(conversation)
+
+                metadata = {}
+                if include_metadata:
+                    metadata = {
+                        "title": conversation.get("title", "ChatGPT Conversation"),
+                        "timestamp": conversation.get("timestamp", "Unknown"),
+                        "message_count": len(conversation.get("messages", [])),
+                        "source": "ChatGPT Export",
+                    }
+
+                doc = Document(text=doc_content, metadata=metadata)
+                docs.append(doc)
+                count += 1
+
+            else:
+                # Create separate documents for each message
+                for message in conversation.get("messages", []):
+                    if max_count > 0 and count >= max_count:
+                        break
+
+                    role = message.get("role", "mixed")
+                    content = message.get("content", "")
+                    msg_timestamp = message.get("timestamp", "")
+
+                    if not content.strip():
+                        continue
+
+                    # Create document content with context
+                    doc_content = f"""Conversation: {conversation.get("title", "ChatGPT Conversation")}
+Role: {role}
+Timestamp: {msg_timestamp or conversation.get("timestamp", "Unknown")}
+Message: {content}
+"""
+
+                    metadata = {}
+                    if include_metadata:
+                        metadata = {
+                            "conversation_title": conversation.get("title", "ChatGPT Conversation"),
+                            "role": role,
+                            "timestamp": msg_timestamp or conversation.get("timestamp", "Unknown"),
+                            "source": "ChatGPT Export",
+                        }
+
+                    doc = Document(text=doc_content, metadata=metadata)
+                    docs.append(doc)
+                    count += 1
+
+        print(f"Created {len(docs)} documents from ChatGPT export")
+        return docs

apps/chatgpt_rag.py

@@ -0,0 +1,186 @@
+"""
+ChatGPT RAG example using the unified interface.
+Supports ChatGPT export data from chat.html files.
+"""
+
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+from base_rag_example import BaseRAGExample
+from chunking import create_text_chunks
+
+from .chatgpt_data.chatgpt_reader import ChatGPTReader
+
+
+class ChatGPTRAG(BaseRAGExample):
+    """RAG example for ChatGPT conversation data."""
+
+    def __init__(self):
+        # Set default values BEFORE calling super().__init__
+        self.max_items_default = -1  # Process all conversations by default
+        self.embedding_model_default = (
+            "sentence-transformers/all-MiniLM-L6-v2"  # Fast 384-dim model
+        )
+
+        super().__init__(
+            name="ChatGPT",
+            description="Process and query ChatGPT conversation exports with LEANN",
+            default_index_name="chatgpt_conversations_index",
+        )
+
+    def _add_specific_arguments(self, parser):
+        """Add ChatGPT-specific arguments."""
+        chatgpt_group = parser.add_argument_group("ChatGPT Parameters")
+        chatgpt_group.add_argument(
+            "--export-path",
+            type=str,
+            default="./chatgpt_export",
+            help="Path to ChatGPT export file (.zip or .html) or directory containing exports (default: ./chatgpt_export)",
+        )
+        chatgpt_group.add_argument(
+            "--concatenate-conversations",
+            action="store_true",
+            default=True,
+            help="Concatenate messages within conversations for better context (default: True)",
+        )
+        chatgpt_group.add_argument(
+            "--separate-messages",
+            action="store_true",
+            help="Process each message as a separate document (overrides --concatenate-conversations)",
+        )
+        chatgpt_group.add_argument(
+            "--chunk-size", type=int, default=512, help="Text chunk size (default: 512)"
+        )
+        chatgpt_group.add_argument(
+            "--chunk-overlap", type=int, default=128, help="Text chunk overlap (default: 128)"
+        )
+
+    def _find_chatgpt_exports(self, export_path: Path) -> list[Path]:
+        """
+        Find ChatGPT export files in the given path.
+
+        Args:
+            export_path: Path to search for exports
+
+        Returns:
+            List of paths to ChatGPT export files
+        """
+        export_files = []
+
+        if export_path.is_file():
+            if export_path.suffix.lower() in [".zip", ".html"]:
+                export_files.append(export_path)
+        elif export_path.is_dir():
+            # Look for zip and html files
+            export_files.extend(export_path.glob("*.zip"))
+            export_files.extend(export_path.glob("*.html"))
+
+        return export_files
+
+    async def load_data(self, args) -> list[str]:
+        """Load ChatGPT export data and convert to text chunks."""
+        export_path = Path(args.export_path)
+
+        if not export_path.exists():
+            print(f"ChatGPT export path not found: {export_path}")
+            print(
+                "Please ensure you have exported your ChatGPT data and placed it in the correct location."
+            )
+            print("\nTo export your ChatGPT data:")
+            print("1. Sign in to ChatGPT")
+            print("2. Click on your profile icon → Settings → Data Controls")
+            print("3. Click 'Export' under Export Data")
+            print("4. Download the zip file from the email link")
+            print("5. Extract or place the file/directory at the specified path")
+            return []
+
+        # Find export files
+        export_files = self._find_chatgpt_exports(export_path)
+
+        if not export_files:
+            print(f"No ChatGPT export files (.zip or .html) found in: {export_path}")
+            return []
+
+        print(f"Found {len(export_files)} ChatGPT export files")
+
+        # Create reader with appropriate settings
+        concatenate = args.concatenate_conversations and not args.separate_messages
+        reader = ChatGPTReader(concatenate_conversations=concatenate)
+
+        # Process each export file
+        all_documents = []
+        total_processed = 0
+
+        for i, export_file in enumerate(export_files):
+            print(f"\nProcessing export file {i + 1}/{len(export_files)}: {export_file.name}")
+
+            try:
+                # Apply max_items limit per file
+                max_per_file = -1
+                if args.max_items > 0:
+                    remaining = args.max_items - total_processed
+                    if remaining <= 0:
+                        break
+                    max_per_file = remaining
+
+                # Load conversations
+                documents = reader.load_data(
+                    chatgpt_export_path=str(export_file),
+                    max_count=max_per_file,
+                    include_metadata=True,
+                )
+
+                if documents:
+                    all_documents.extend(documents)
+                    total_processed += len(documents)
+                    print(f"Processed {len(documents)} conversations from this file")
+                else:
+                    print(f"No conversations loaded from {export_file}")
+
+            except Exception as e:
+                print(f"Error processing {export_file}: {e}")
+                continue
+
+        if not all_documents:
+            print("No conversations found to process!")
+            print("\nTroubleshooting:")
+            print("- Ensure the export file is a valid ChatGPT export")
+            print("- Check that the HTML file contains conversation data")
+            print("- Try extracting the zip file and pointing to the HTML file directly")
+            return []
+
+        print(f"\nTotal conversations processed: {len(all_documents)}")
+        print("Now starting to split into text chunks... this may take some time")
+
+        # Convert to text chunks
+        all_texts = create_text_chunks(
+            all_documents, chunk_size=args.chunk_size, chunk_overlap=args.chunk_overlap
+        )
+
+        print(f"Created {len(all_texts)} text chunks from {len(all_documents)} conversations")
+        return all_texts
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Example queries for ChatGPT RAG
+    print("\n🤖 ChatGPT RAG Example")
+    print("=" * 50)
+    print("\nExample queries you can try:")
+    print("- 'What did I ask about Python programming?'")
+    print("- 'Show me conversations about machine learning'")
+    print("- 'Find discussions about travel planning'")
+    print("- 'What advice did ChatGPT give me about career development?'")
+    print("- 'Search for conversations about cooking recipes'")
+    print("\nTo get started:")
+    print("1. Export your ChatGPT data from Settings → Data Controls → Export")
+    print("2. Place the downloaded zip file or extracted HTML in ./chatgpt_export/")
+    print("3. Run this script to build your personal ChatGPT knowledge base!")
+    print("\nOr run without --query for interactive mode\n")
+
+    rag = ChatGPTRAG()
+    asyncio.run(rag.run())

apps/chunking/__init__.py

@@ -0,0 +1,44 @@
+"""Unified chunking utilities facade.
+
+This module re-exports the packaged utilities from `leann.chunking_utils` so
+that both repo apps (importing `chunking`) and installed wheels share one
+single implementation. When running from the repo without installation, it
+adds the `packages/leann-core/src` directory to `sys.path` as a fallback.
+"""
+
+import sys
+from pathlib import Path
+
+try:
+    from leann.chunking_utils import (
+        CODE_EXTENSIONS,
+        create_ast_chunks,
+        create_text_chunks,
+        create_traditional_chunks,
+        detect_code_files,
+        get_language_from_extension,
+    )
+except Exception:  # pragma: no cover - best-effort fallback for dev environment
+    repo_root = Path(__file__).resolve().parents[2]
+    leann_src = repo_root / "packages" / "leann-core" / "src"
+    if leann_src.exists():
+        sys.path.insert(0, str(leann_src))
+        from leann.chunking_utils import (
+            CODE_EXTENSIONS,
+            create_ast_chunks,
+            create_text_chunks,
+            create_traditional_chunks,
+            detect_code_files,
+            get_language_from_extension,
+        )
+    else:
+        raise
+
+__all__ = [
+    "CODE_EXTENSIONS",
+    "create_ast_chunks",
+    "create_text_chunks",
+    "create_traditional_chunks",
+    "detect_code_files",
+    "get_language_from_extension",
+]

apps/claude_data/claude_reader.py

@@ -0,0 +1,420 @@
+"""
+Claude export data reader.
+
+Reads and processes Claude conversation data from exported JSON files.
+"""
+
+import json
+from pathlib import Path
+from typing import Any
+from zipfile import ZipFile
+
+from llama_index.core import Document
+from llama_index.core.readers.base import BaseReader
+
+
+class ClaudeReader(BaseReader):
+    """
+    Claude export data reader.
+
+    Reads Claude conversation data from exported JSON files or zip archives.
+    Processes conversations into structured documents with metadata.
+    """
+
+    def __init__(self, concatenate_conversations: bool = True) -> None:
+        """
+        Initialize.
+
+        Args:
+            concatenate_conversations: Whether to concatenate messages within conversations for better context
+        """
+        self.concatenate_conversations = concatenate_conversations
+
+    def _extract_json_from_zip(self, zip_path: Path) -> list[str]:
+        """
+        Extract JSON files from Claude export zip file.
+
+        Args:
+            zip_path: Path to the Claude export zip file
+
+        Returns:
+            List of JSON content strings, or empty list if not found
+        """
+        json_contents = []
+        try:
+            with ZipFile(zip_path, "r") as zip_file:
+                # Look for JSON files
+                json_files = [f for f in zip_file.namelist() if f.endswith(".json")]
+
+                if not json_files:
+                    print(f"No JSON files found in {zip_path}")
+                    return []
+
+                print(f"Found {len(json_files)} JSON files in archive")
+
+                for json_file in json_files:
+                    with zip_file.open(json_file) as f:
+                        content = f.read().decode("utf-8", errors="ignore")
+                        json_contents.append(content)
+
+        except Exception as e:
+            print(f"Error extracting JSON from zip {zip_path}: {e}")
+
+        return json_contents
+
+    def _parse_claude_json(self, json_content: str) -> list[dict]:
+        """
+        Parse Claude JSON export to extract conversations.
+
+        Args:
+            json_content: JSON content from Claude export
+
+        Returns:
+            List of conversation dictionaries
+        """
+        try:
+            data = json.loads(json_content)
+        except json.JSONDecodeError as e:
+            print(f"Error parsing JSON: {e}")
+            return []
+
+        conversations = []
+
+        # Handle different possible JSON structures
+        if isinstance(data, list):
+            # If data is a list of conversations
+            for item in data:
+                conversation = self._extract_conversation_from_json(item)
+                if conversation:
+                    conversations.append(conversation)
+        elif isinstance(data, dict):
+            # Check for common structures
+            if "conversations" in data:
+                # Structure: {"conversations": [...]}
+                for item in data["conversations"]:
+                    conversation = self._extract_conversation_from_json(item)
+                    if conversation:
+                        conversations.append(conversation)
+            elif "messages" in data:
+                # Single conversation with messages
+                conversation = self._extract_conversation_from_json(data)
+                if conversation:
+                    conversations.append(conversation)
+            else:
+                # Try to treat the whole object as a conversation
+                conversation = self._extract_conversation_from_json(data)
+                if conversation:
+                    conversations.append(conversation)
+
+        return conversations
+
+    def _extract_conversation_from_json(self, conv_data: dict) -> dict | None:
+        """
+        Extract conversation data from a JSON object.
+
+        Args:
+            conv_data: Dictionary containing conversation data
+
+        Returns:
+            Dictionary with conversation data or None
+        """
+        if not isinstance(conv_data, dict):
+            return None
+
+        messages = []
+
+        # Look for messages in various possible structures
+        message_sources = []
+        if "messages" in conv_data:
+            message_sources = conv_data["messages"]
+        elif "chat" in conv_data:
+            message_sources = conv_data["chat"]
+        elif "conversation" in conv_data:
+            message_sources = conv_data["conversation"]
+        else:
+            # If no clear message structure, try to extract from the object itself
+            if "content" in conv_data and "role" in conv_data:
+                message_sources = [conv_data]
+
+        for msg_data in message_sources:
+            message = self._extract_message_from_json(msg_data)
+            if message:
+                messages.append(message)
+
+        if not messages:
+            return None
+
+        # Extract conversation metadata
+        title = self._extract_title_from_conversation(conv_data, messages)
+        timestamp = self._extract_timestamp_from_conversation(conv_data)
+
+        return {"title": title, "messages": messages, "timestamp": timestamp}
+
+    def _extract_message_from_json(self, msg_data: dict) -> dict | None:
+        """
+        Extract message data from a JSON message object.
+
+        Args:
+            msg_data: Dictionary containing message data
+
+        Returns:
+            Dictionary with message data or None
+        """
+        if not isinstance(msg_data, dict):
+            return None
+
+        # Extract content from various possible fields
+        content = ""
+        content_fields = ["content", "text", "message", "body"]
+        for field in content_fields:
+            if msg_data.get(field):
+                content = str(msg_data[field])
+                break
+
+        if not content or len(content.strip()) < 3:
+            return None
+
+        # Extract role (user/assistant/human/ai/claude)
+        role = "mixed"  # Default role
+        role_fields = ["role", "sender", "from", "author", "type"]
+        for field in role_fields:
+            if msg_data.get(field):
+                role_value = str(msg_data[field]).lower()
+                if role_value in ["user", "human", "person"]:
+                    role = "user"
+                elif role_value in ["assistant", "ai", "claude", "bot"]:
+                    role = "assistant"
+                break
+
+        # Extract timestamp
+        timestamp = self._extract_timestamp_from_message(msg_data)
+
+        return {"role": role, "content": content, "timestamp": timestamp}
+
+    def _extract_timestamp_from_message(self, msg_data: dict) -> str | None:
+        """Extract timestamp from message data."""
+        timestamp_fields = ["timestamp", "created_at", "date", "time"]
+        for field in timestamp_fields:
+            if msg_data.get(field):
+                return str(msg_data[field])
+        return None
+
+    def _extract_timestamp_from_conversation(self, conv_data: dict) -> str | None:
+        """Extract timestamp from conversation data."""
+        timestamp_fields = ["timestamp", "created_at", "date", "updated_at", "last_updated"]
+        for field in timestamp_fields:
+            if conv_data.get(field):
+                return str(conv_data[field])
+        return None
+
+    def _extract_title_from_conversation(self, conv_data: dict, messages: list) -> str:
+        """Extract or generate title for conversation."""
+        # Try to find explicit title
+        title_fields = ["title", "name", "subject", "topic"]
+        for field in title_fields:
+            if conv_data.get(field):
+                return str(conv_data[field])
+
+        # Generate title from first user message
+        for message in messages:
+            if message.get("role") == "user":
+                content = message.get("content", "")
+                if content:
+                    # Use first 50 characters as title
+                    title = content[:50].strip()
+                    if len(content) > 50:
+                        title += "..."
+                    return title
+
+        return "Claude Conversation"
+
+    def _create_concatenated_content(self, conversation: dict) -> str:
+        """
+        Create concatenated content from conversation messages.
+
+        Args:
+            conversation: Dictionary containing conversation data
+
+        Returns:
+            Formatted concatenated content
+        """
+        title = conversation.get("title", "Claude Conversation")
+        messages = conversation.get("messages", [])
+        timestamp = conversation.get("timestamp", "Unknown")
+
+        # Build message content
+        message_parts = []
+        for message in messages:
+            role = message.get("role", "mixed")
+            content = message.get("content", "")
+            msg_timestamp = message.get("timestamp", "")
+
+            if role == "user":
+                prefix = "[You]"
+            elif role == "assistant":
+                prefix = "[Claude]"
+            else:
+                prefix = "[Message]"
+
+            # Add timestamp if available
+            if msg_timestamp:
+                prefix += f" ({msg_timestamp})"
+
+            message_parts.append(f"{prefix}: {content}")
+
+        concatenated_text = "\n\n".join(message_parts)
+
+        # Create final document content
+        doc_content = f"""Conversation: {title}
+Date: {timestamp}
+Messages ({len(messages)} messages):
+
+{concatenated_text}
+"""
+        return doc_content
+
+    def load_data(self, input_dir: str | None = None, **load_kwargs: Any) -> list[Document]:
+        """
+        Load Claude export data.
+
+        Args:
+            input_dir: Directory containing Claude export files or path to specific file
+            **load_kwargs:
+                max_count (int): Maximum number of conversations to process
+                claude_export_path (str): Specific path to Claude export file/directory
+                include_metadata (bool): Whether to include metadata in documents
+        """
+        docs: list[Document] = []
+        max_count = load_kwargs.get("max_count", -1)
+        claude_export_path = load_kwargs.get("claude_export_path", input_dir)
+        include_metadata = load_kwargs.get("include_metadata", True)
+
+        if not claude_export_path:
+            print("No Claude export path provided")
+            return docs
+
+        export_path = Path(claude_export_path)
+
+        if not export_path.exists():
+            print(f"Claude export path not found: {export_path}")
+            return docs
+
+        json_contents = []
+
+        # Handle different input types
+        if export_path.is_file():
+            if export_path.suffix.lower() == ".zip":
+                # Extract JSON from zip file
+                json_contents = self._extract_json_from_zip(export_path)
+            elif export_path.suffix.lower() == ".json":
+                # Read JSON file directly
+                try:
+                    with open(export_path, encoding="utf-8", errors="ignore") as f:
+                        json_contents.append(f.read())
+                except Exception as e:
+                    print(f"Error reading JSON file {export_path}: {e}")
+                    return docs
+            else:
+                print(f"Unsupported file type: {export_path.suffix}")
+                return docs
+
+        elif export_path.is_dir():
+            # Look for JSON files in directory
+            json_files = list(export_path.glob("*.json"))
+            zip_files = list(export_path.glob("*.zip"))
+
+            if json_files:
+                print(f"Found {len(json_files)} JSON files in directory")
+                for json_file in json_files:
+                    try:
+                        with open(json_file, encoding="utf-8", errors="ignore") as f:
+                            json_contents.append(f.read())
+                    except Exception as e:
+                        print(f"Error reading JSON file {json_file}: {e}")
+                        continue
+
+            if zip_files:
+                print(f"Found {len(zip_files)} ZIP files in directory")
+                for zip_file in zip_files:
+                    zip_contents = self._extract_json_from_zip(zip_file)
+                    json_contents.extend(zip_contents)
+
+            if not json_files and not zip_files:
+                print(f"No JSON or ZIP files found in {export_path}")
+                return docs
+
+        if not json_contents:
+            print("No JSON content found to process")
+            return docs
+
+        # Parse conversations from JSON content
+        print("Parsing Claude conversations from JSON...")
+        all_conversations = []
+        for json_content in json_contents:
+            conversations = self._parse_claude_json(json_content)
+            all_conversations.extend(conversations)
+
+        if not all_conversations:
+            print("No conversations found in JSON content")
+            return docs
+
+        print(f"Found {len(all_conversations)} conversations")
+
+        # Process conversations into documents
+        count = 0
+        for conversation in all_conversations:
+            if max_count > 0 and count >= max_count:
+                break
+
+            if self.concatenate_conversations:
+                # Create one document per conversation with concatenated messages
+                doc_content = self._create_concatenated_content(conversation)
+
+                metadata = {}
+                if include_metadata:
+                    metadata = {
+                        "title": conversation.get("title", "Claude Conversation"),
+                        "timestamp": conversation.get("timestamp", "Unknown"),
+                        "message_count": len(conversation.get("messages", [])),
+                        "source": "Claude Export",
+                    }
+
+                doc = Document(text=doc_content, metadata=metadata)
+                docs.append(doc)
+                count += 1
+
+            else:
+                # Create separate documents for each message
+                for message in conversation.get("messages", []):
+                    if max_count > 0 and count >= max_count:
+                        break
+
+                    role = message.get("role", "mixed")
+                    content = message.get("content", "")
+                    msg_timestamp = message.get("timestamp", "")
+
+                    if not content.strip():
+                        continue
+
+                    # Create document content with context
+                    doc_content = f"""Conversation: {conversation.get("title", "Claude Conversation")}
+Role: {role}
+Timestamp: {msg_timestamp or conversation.get("timestamp", "Unknown")}
+Message: {content}
+"""
+
+                    metadata = {}
+                    if include_metadata:
+                        metadata = {
+                            "conversation_title": conversation.get("title", "Claude Conversation"),
+                            "role": role,
+                            "timestamp": msg_timestamp or conversation.get("timestamp", "Unknown"),
+                            "source": "Claude Export",
+                        }
+
+                    doc = Document(text=doc_content, metadata=metadata)
+                    docs.append(doc)
+                    count += 1
+
+        print(f"Created {len(docs)} documents from Claude export")
+        return docs

apps/claude_rag.py

@@ -0,0 +1,189 @@
+"""
+Claude RAG example using the unified interface.
+Supports Claude export data from JSON files.
+"""
+
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+from base_rag_example import BaseRAGExample
+from chunking import create_text_chunks
+
+from .claude_data.claude_reader import ClaudeReader
+
+
+class ClaudeRAG(BaseRAGExample):
+    """RAG example for Claude conversation data."""
+
+    def __init__(self):
+        # Set default values BEFORE calling super().__init__
+        self.max_items_default = -1  # Process all conversations by default
+        self.embedding_model_default = (
+            "sentence-transformers/all-MiniLM-L6-v2"  # Fast 384-dim model
+        )
+
+        super().__init__(
+            name="Claude",
+            description="Process and query Claude conversation exports with LEANN",
+            default_index_name="claude_conversations_index",
+        )
+
+    def _add_specific_arguments(self, parser):
+        """Add Claude-specific arguments."""
+        claude_group = parser.add_argument_group("Claude Parameters")
+        claude_group.add_argument(
+            "--export-path",
+            type=str,
+            default="./claude_export",
+            help="Path to Claude export file (.json or .zip) or directory containing exports (default: ./claude_export)",
+        )
+        claude_group.add_argument(
+            "--concatenate-conversations",
+            action="store_true",
+            default=True,
+            help="Concatenate messages within conversations for better context (default: True)",
+        )
+        claude_group.add_argument(
+            "--separate-messages",
+            action="store_true",
+            help="Process each message as a separate document (overrides --concatenate-conversations)",
+        )
+        claude_group.add_argument(
+            "--chunk-size", type=int, default=512, help="Text chunk size (default: 512)"
+        )
+        claude_group.add_argument(
+            "--chunk-overlap", type=int, default=128, help="Text chunk overlap (default: 128)"
+        )
+
+    def _find_claude_exports(self, export_path: Path) -> list[Path]:
+        """
+        Find Claude export files in the given path.
+
+        Args:
+            export_path: Path to search for exports
+
+        Returns:
+            List of paths to Claude export files
+        """
+        export_files = []
+
+        if export_path.is_file():
+            if export_path.suffix.lower() in [".zip", ".json"]:
+                export_files.append(export_path)
+        elif export_path.is_dir():
+            # Look for zip and json files
+            export_files.extend(export_path.glob("*.zip"))
+            export_files.extend(export_path.glob("*.json"))
+
+        return export_files
+
+    async def load_data(self, args) -> list[str]:
+        """Load Claude export data and convert to text chunks."""
+        export_path = Path(args.export_path)
+
+        if not export_path.exists():
+            print(f"Claude export path not found: {export_path}")
+            print(
+                "Please ensure you have exported your Claude data and placed it in the correct location."
+            )
+            print("\nTo export your Claude data:")
+            print("1. Open Claude in your browser")
+            print("2. Look for export/download options in settings or conversation menu")
+            print("3. Download the conversation data (usually in JSON format)")
+            print("4. Place the file/directory at the specified path")
+            print(
+                "\nNote: Claude export methods may vary. Check Claude's help documentation for current instructions."
+            )
+            return []
+
+        # Find export files
+        export_files = self._find_claude_exports(export_path)
+
+        if not export_files:
+            print(f"No Claude export files (.json or .zip) found in: {export_path}")
+            return []
+
+        print(f"Found {len(export_files)} Claude export files")
+
+        # Create reader with appropriate settings
+        concatenate = args.concatenate_conversations and not args.separate_messages
+        reader = ClaudeReader(concatenate_conversations=concatenate)
+
+        # Process each export file
+        all_documents = []
+        total_processed = 0
+
+        for i, export_file in enumerate(export_files):
+            print(f"\nProcessing export file {i + 1}/{len(export_files)}: {export_file.name}")
+
+            try:
+                # Apply max_items limit per file
+                max_per_file = -1
+                if args.max_items > 0:
+                    remaining = args.max_items - total_processed
+                    if remaining <= 0:
+                        break
+                    max_per_file = remaining
+
+                # Load conversations
+                documents = reader.load_data(
+                    claude_export_path=str(export_file),
+                    max_count=max_per_file,
+                    include_metadata=True,
+                )
+
+                if documents:
+                    all_documents.extend(documents)
+                    total_processed += len(documents)
+                    print(f"Processed {len(documents)} conversations from this file")
+                else:
+                    print(f"No conversations loaded from {export_file}")
+
+            except Exception as e:
+                print(f"Error processing {export_file}: {e}")
+                continue
+
+        if not all_documents:
+            print("No conversations found to process!")
+            print("\nTroubleshooting:")
+            print("- Ensure the export file is a valid Claude export")
+            print("- Check that the JSON file contains conversation data")
+            print("- Try using a different export format or method")
+            print("- Check Claude's documentation for current export procedures")
+            return []
+
+        print(f"\nTotal conversations processed: {len(all_documents)}")
+        print("Now starting to split into text chunks... this may take some time")
+
+        # Convert to text chunks
+        all_texts = create_text_chunks(
+            all_documents, chunk_size=args.chunk_size, chunk_overlap=args.chunk_overlap
+        )
+
+        print(f"Created {len(all_texts)} text chunks from {len(all_documents)} conversations")
+        return all_texts
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Example queries for Claude RAG
+    print("\n🤖 Claude RAG Example")
+    print("=" * 50)
+    print("\nExample queries you can try:")
+    print("- 'What did I ask Claude about Python programming?'")
+    print("- 'Show me conversations about machine learning'")
+    print("- 'Find discussions about code optimization'")
+    print("- 'What advice did Claude give me about software design?'")
+    print("- 'Search for conversations about debugging techniques'")
+    print("\nTo get started:")
+    print("1. Export your Claude conversation data")
+    print("2. Place the JSON/ZIP file in ./claude_export/")
+    print("3. Run this script to build your personal Claude knowledge base!")
+    print("\nOr run without --query for interactive mode\n")
+
+    rag = ClaudeRAG()
+    asyncio.run(rag.run())

apps/code_rag.py

@@ -0,0 +1,211 @@
+"""
+Code RAG example using AST-aware chunking for optimal code understanding.
+Specialized for code repositories with automatic language detection and
+optimized chunking parameters.
+"""
+
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+from base_rag_example import BaseRAGExample
+from chunking import CODE_EXTENSIONS, create_text_chunks
+from llama_index.core import SimpleDirectoryReader
+
+
+class CodeRAG(BaseRAGExample):
+    """Specialized RAG example for code repositories with AST-aware chunking."""
+
+    def __init__(self):
+        super().__init__(
+            name="Code",
+            description="Process and query code repositories with AST-aware chunking",
+            default_index_name="code_index",
+        )
+        # Override defaults for code-specific usage
+        self.embedding_model_default = "facebook/contriever"  # Good for code
+        self.max_items_default = -1  # Process all code files by default
+
+    def _add_specific_arguments(self, parser):
+        """Add code-specific arguments."""
+        code_group = parser.add_argument_group("Code Repository Parameters")
+
+        code_group.add_argument(
+            "--repo-dir",
+            type=str,
+            default=".",
+            help="Code repository directory to index (default: current directory)",
+        )
+        code_group.add_argument(
+            "--include-extensions",
+            nargs="+",
+            default=list(CODE_EXTENSIONS.keys()),
+            help="File extensions to include (default: supported code extensions)",
+        )
+        code_group.add_argument(
+            "--exclude-dirs",
+            nargs="+",
+            default=[
+                ".git",
+                "__pycache__",
+                "node_modules",
+                "venv",
+                ".venv",
+                "build",
+                "dist",
+                "target",
+            ],
+            help="Directories to exclude from indexing",
+        )
+        code_group.add_argument(
+            "--max-file-size",
+            type=int,
+            default=1000000,  # 1MB
+            help="Maximum file size in bytes to process (default: 1MB)",
+        )
+        code_group.add_argument(
+            "--include-comments",
+            action="store_true",
+            help="Include comments in chunking (useful for documentation)",
+        )
+        code_group.add_argument(
+            "--preserve-imports",
+            action="store_true",
+            default=True,
+            help="Try to preserve import statements in chunks (default: True)",
+        )
+
+    async def load_data(self, args) -> list[str]:
+        """Load code files and convert to AST-aware chunks."""
+        print(f"🔍 Scanning code repository: {args.repo_dir}")
+        print(f"📁 Including extensions: {args.include_extensions}")
+        print(f"🚫 Excluding directories: {args.exclude_dirs}")
+
+        # Check if repository directory exists
+        repo_path = Path(args.repo_dir)
+        if not repo_path.exists():
+            raise ValueError(f"Repository directory not found: {args.repo_dir}")
+
+        # Load code files with filtering
+        reader_kwargs = {
+            "recursive": True,
+            "encoding": "utf-8",
+            "required_exts": args.include_extensions,
+            "exclude_hidden": True,
+        }
+
+        # Create exclusion filter
+        def file_filter(file_path: str) -> bool:
+            """Filter out unwanted files and directories."""
+            path = Path(file_path)
+
+            # Check file size
+            try:
+                if path.stat().st_size > args.max_file_size:
+                    print(f"⚠️ Skipping large file: {path.name} ({path.stat().st_size} bytes)")
+                    return False
+            except Exception:
+                return False
+
+            # Check if in excluded directory
+            for exclude_dir in args.exclude_dirs:
+                if exclude_dir in path.parts:
+                    return False
+
+            return True
+
+        try:
+            # Load documents with file filtering
+            documents = SimpleDirectoryReader(
+                args.repo_dir,
+                file_extractor=None,  # Use default extractors
+                **reader_kwargs,
+            ).load_data(show_progress=True)
+
+            # Apply custom filtering
+            filtered_docs = []
+            for doc in documents:
+                file_path = doc.metadata.get("file_path", "")
+                if file_filter(file_path):
+                    filtered_docs.append(doc)
+
+            documents = filtered_docs
+
+        except Exception as e:
+            print(f"❌ Error loading code files: {e}")
+            return []
+
+        if not documents:
+            print(
+                f"❌ No code files found in {args.repo_dir} with extensions {args.include_extensions}"
+            )
+            return []
+
+        print(f"✅ Loaded {len(documents)} code files")
+
+        # Show breakdown by language/extension
+        ext_counts = {}
+        for doc in documents:
+            file_path = doc.metadata.get("file_path", "")
+            if file_path:
+                ext = Path(file_path).suffix.lower()
+                ext_counts[ext] = ext_counts.get(ext, 0) + 1
+
+        print("📊 Files by extension:")
+        for ext, count in sorted(ext_counts.items()):
+            print(f"   {ext}: {count} files")
+
+        # Use AST-aware chunking by default for code
+        print(
+            f"🧠 Using AST-aware chunking (chunk_size: {args.ast_chunk_size}, overlap: {args.ast_chunk_overlap})"
+        )
+
+        all_texts = create_text_chunks(
+            documents,
+            chunk_size=256,  # Fallback for non-code files
+            chunk_overlap=64,
+            use_ast_chunking=True,  # Always use AST for code RAG
+            ast_chunk_size=args.ast_chunk_size,
+            ast_chunk_overlap=args.ast_chunk_overlap,
+            code_file_extensions=args.include_extensions,
+            ast_fallback_traditional=True,
+        )
+
+        # Apply max_items limit if specified
+        if args.max_items > 0 and len(all_texts) > args.max_items:
+            print(f"⏳ Limiting to {args.max_items} chunks (from {len(all_texts)})")
+            all_texts = all_texts[: args.max_items]
+
+        print(f"✅ Generated {len(all_texts)} code chunks")
+        return all_texts
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Example queries for code RAG
+    print("\n💻 Code RAG Example")
+    print("=" * 50)
+    print("\nExample queries you can try:")
+    print("- 'How does the embedding computation work?'")
+    print("- 'What are the main classes in this codebase?'")
+    print("- 'Show me the search implementation'")
+    print("- 'How is error handling implemented?'")
+    print("- 'What design patterns are used?'")
+    print("- 'Explain the chunking logic'")
+    print("\n🚀 Features:")
+    print("- ✅ AST-aware chunking preserves code structure")
+    print("- ✅ Automatic language detection")
+    print("- ✅ Smart filtering of large files and common excludes")
+    print("- ✅ Optimized for code understanding")
+    print("\nUsage examples:")
+    print("  python -m apps.code_rag --repo-dir ./my_project")
+    print(
+        "  python -m apps.code_rag --include-extensions .py .js --query 'How does authentication work?'"
+    )
+    print("\nOr run without --query for interactive mode\n")
+
+    rag = CodeRAG()
+    asyncio.run(rag.run())

apps/document_rag.py

@@ -0,0 +1,131 @@
+"""
+Document RAG example using the unified interface.
+Supports PDF, TXT, MD, and other document formats.
+"""
+
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+from base_rag_example import BaseRAGExample
+from chunking import create_text_chunks
+from llama_index.core import SimpleDirectoryReader
+
+
+class DocumentRAG(BaseRAGExample):
+    """RAG example for document processing (PDF, TXT, MD, etc.)."""
+
+    def __init__(self):
+        super().__init__(
+            name="Document",
+            description="Process and query documents (PDF, TXT, MD, etc.) with LEANN",
+            default_index_name="test_doc_files",
+        )
+
+    def _add_specific_arguments(self, parser):
+        """Add document-specific arguments."""
+        doc_group = parser.add_argument_group("Document Parameters")
+        doc_group.add_argument(
+            "--data-dir",
+            type=str,
+            default="data",
+            help="Directory containing documents to index (default: data)",
+        )
+        doc_group.add_argument(
+            "--file-types",
+            nargs="+",
+            default=None,
+            help="Filter by file types (e.g., .pdf .txt .md). If not specified, all supported types are processed",
+        )
+        doc_group.add_argument(
+            "--chunk-size", type=int, default=256, help="Text chunk size (default: 256)"
+        )
+        doc_group.add_argument(
+            "--chunk-overlap", type=int, default=128, help="Text chunk overlap (default: 128)"
+        )
+        doc_group.add_argument(
+            "--enable-code-chunking",
+            action="store_true",
+            help="Enable AST-aware chunking for code files in the data directory",
+        )
+
+    async def load_data(self, args) -> list[str]:
+        """Load documents and convert to text chunks."""
+        print(f"Loading documents from: {args.data_dir}")
+        if args.file_types:
+            print(f"Filtering by file types: {args.file_types}")
+        else:
+            print("Processing all supported file types")
+
+        # Check if data directory exists
+        data_path = Path(args.data_dir)
+        if not data_path.exists():
+            raise ValueError(f"Data directory not found: {args.data_dir}")
+
+        # Load documents
+        reader_kwargs = {
+            "recursive": True,
+            "encoding": "utf-8",
+        }
+        if args.file_types:
+            reader_kwargs["required_exts"] = args.file_types
+
+        documents = SimpleDirectoryReader(args.data_dir, **reader_kwargs).load_data(
+            show_progress=True
+        )
+
+        if not documents:
+            print(f"No documents found in {args.data_dir} with extensions {args.file_types}")
+            return []
+
+        print(f"Loaded {len(documents)} documents")
+
+        # Determine chunking strategy
+        use_ast = args.enable_code_chunking or getattr(args, "use_ast_chunking", False)
+
+        if use_ast:
+            print("Using AST-aware chunking for code files")
+
+        # Convert to text chunks with optional AST support
+        all_texts = create_text_chunks(
+            documents,
+            chunk_size=args.chunk_size,
+            chunk_overlap=args.chunk_overlap,
+            use_ast_chunking=use_ast,
+            ast_chunk_size=getattr(args, "ast_chunk_size", 512),
+            ast_chunk_overlap=getattr(args, "ast_chunk_overlap", 64),
+            code_file_extensions=getattr(args, "code_file_extensions", None),
+            ast_fallback_traditional=getattr(args, "ast_fallback_traditional", True),
+        )
+
+        # Apply max_items limit if specified
+        if args.max_items > 0 and len(all_texts) > args.max_items:
+            print(f"Limiting to {args.max_items} chunks (from {len(all_texts)})")
+            all_texts = all_texts[: args.max_items]
+
+        return all_texts
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Example queries for document RAG
+    print("\n📄 Document RAG Example")
+    print("=" * 50)
+    print("\nExample queries you can try:")
+    print("- 'What are the main techniques LEANN uses?'")
+    print("- 'What is the technique DLPM?'")
+    print("- 'Who does Elizabeth Bennet marry?'")
+    print(
+        "- 'What is the problem of developing pan gu model Huawei meets? (盘古大模型开发中遇到什么问题?)'"
+    )
+    print("\n🚀 NEW: Code-aware chunking available!")
+    print("- Use --enable-code-chunking to enable AST-aware chunking for code files")
+    print("- Supports Python, Java, C#, TypeScript files")
+    print("- Better semantic understanding of code structure")
+    print("\nOr run without --query for interactive mode\n")
+
+    rag = DocumentRAG()
+    asyncio.run(rag.run())

apps/email_data/LEANN_email_reader.py

@@ -0,0 +1,167 @@
+import email
+import os
+from pathlib import Path
+from typing import Any
+
+from llama_index.core import Document
+from llama_index.core.readers.base import BaseReader
+
+
+def find_all_messages_directories(root: str | None = None) -> list[Path]:
+    """
+    Recursively find all 'Messages' directories under the given root.
+    Returns a list of Path objects.
+    """
+    if root is None:
+        # Auto-detect user's mail path
+        home_dir = os.path.expanduser("~")
+        root = os.path.join(home_dir, "Library", "Mail")
+
+    messages_dirs = []
+    for dirpath, _dirnames, _filenames in os.walk(root):
+        if os.path.basename(dirpath) == "Messages":
+            messages_dirs.append(Path(dirpath))
+    return messages_dirs
+
+
+class EmlxReader(BaseReader):
+    """
+    Apple Mail .emlx file reader with embedded metadata.
+
+    Reads individual .emlx files from Apple Mail's storage format.
+    """
+
+    def __init__(self, include_html: bool = False) -> None:
+        """
+        Initialize.
+
+        Args:
+            include_html: Whether to include HTML content in the email body (default: False)
+        """
+        self.include_html = include_html
+
+    def load_data(self, input_dir: str, **load_kwargs: Any) -> list[Document]:
+        """
+        Load data from the input directory containing .emlx files.
+
+        Args:
+            input_dir: Directory containing .emlx files
+            **load_kwargs:
+                max_count (int): Maximum amount of messages to read.
+        """
+        docs: list[Document] = []
+        max_count = load_kwargs.get("max_count", 1000)
+        count = 0
+        total_files = 0
+        successful_files = 0
+        failed_files = 0
+
+        print(f"Starting to process directory: {input_dir}")
+
+        # Walk through the directory recursively
+        for dirpath, dirnames, filenames in os.walk(input_dir):
+            # Skip hidden directories
+            dirnames[:] = [d for d in dirnames if not d.startswith(".")]
+
+            for filename in filenames:
+                # Check if we've reached the max count (skip if max_count == -1)
+                if max_count > 0 and count >= max_count:
+                    break
+
+                if filename.endswith(".emlx"):
+                    total_files += 1
+                    filepath = os.path.join(dirpath, filename)
+                    try:
+                        # Read the .emlx file
+                        with open(filepath, encoding="utf-8", errors="ignore") as f:
+                            content = f.read()
+
+                        # .emlx files have a length prefix followed by the email content
+                        # The first line contains the length, followed by the email
+                        lines = content.split("\n", 1)
+                        if len(lines) >= 2:
+                            email_content = lines[1]
+
+                            # Parse the email using Python's email module
+                            try:
+                                msg = email.message_from_string(email_content)
+
+                                # Extract email metadata
+                                subject = msg.get("Subject", "No Subject")
+                                from_addr = msg.get("From", "Unknown")
+                                to_addr = msg.get("To", "Unknown")
+                                date = msg.get("Date", "Unknown")
+
+                                # Extract email body
+                                body = ""
+                                if msg.is_multipart():
+                                    for part in msg.walk():
+                                        if (
+                                            part.get_content_type() == "text/plain"
+                                            or part.get_content_type() == "text/html"
+                                        ):
+                                            if (
+                                                part.get_content_type() == "text/html"
+                                                and not self.include_html
+                                            ):
+                                                continue
+                                            try:
+                                                payload = part.get_payload(decode=True)
+                                                if payload:
+                                                    body += payload.decode("utf-8", errors="ignore")
+                                            except Exception as e:
+                                                print(f"Error decoding payload: {e}")
+                                                continue
+                                else:
+                                    try:
+                                        payload = msg.get_payload(decode=True)
+                                        if payload:
+                                            body = payload.decode("utf-8", errors="ignore")
+                                    except Exception as e:
+                                        print(f"Error decoding single part payload: {e}")
+                                        body = ""
+
+                                # Only create document if we have some content
+                                if body.strip() or subject != "No Subject":
+                                    # Create document content with metadata embedded in text
+                                    doc_content = f"""
+[File]: {filename}
+[From]: {from_addr}
+[To]: {to_addr}
+[Subject]: {subject}
+[Date]: {date}
+[EMAIL BODY Start]:
+{body}
+"""
+
+                                    # No separate metadata - everything is in the text
+                                    doc = Document(text=doc_content, metadata={})
+                                    docs.append(doc)
+                                    count += 1
+                                    successful_files += 1
+
+                                    # Print first few successful files for debugging
+                                    if successful_files <= 3:
+                                        print(
+                                            f"Successfully loaded: {filename} - Subject: {subject[:50]}..."
+                                        )
+
+                            except Exception as e:
+                                failed_files += 1
+                                if failed_files <= 5:  # Only print first few errors
+                                    print(f"Error parsing email from {filepath}: {e}")
+                                continue
+
+                    except Exception as e:
+                        failed_files += 1
+                        if failed_files <= 5:  # Only print first few errors
+                            print(f"Error reading file {filepath}: {e}")
+                        continue
+
+        print("Processing summary:")
+        print(f"  Total .emlx files found: {total_files}")
+        print(f"  Successfully loaded: {successful_files}")
+        print(f"  Failed to load: {failed_files}")
+        print(f"  Final documents: {len(docs)}")
+
+        return docs

apps/email_data/email.py

@@ -7,9 +7,9 @@ Contains simple parser for mbox files.
 
 import logging
 from pathlib import Path
-from typing import Any, Dict, List, Optional
-from fsspec import AbstractFileSystem
+from typing import Any
 
+from fsspec import AbstractFileSystem
 from llama_index.core.readers.base import BaseReader
 from llama_index.core.schema import Document
 
@@ -27,11 +27,7 @@ class MboxReader(BaseReader):
     """
 
     DEFAULT_MESSAGE_FORMAT: str = (
-        "Date: {_date}\n"
-        "From: {_from}\n"
-        "To: {_to}\n"
-        "Subject: {_subject}\n"
-        "Content: {_content}"
+        "Date: {_date}\nFrom: {_from}\nTo: {_to}\nSubject: {_subject}\nContent: {_content}"
     )
 
     def __init__(
@@ -45,9 +41,7 @@ class MboxReader(BaseReader):
         try:
             from bs4 import BeautifulSoup  # noqa
         except ImportError:
-            raise ImportError(
-                "`beautifulsoup4` package not found: `pip install beautifulsoup4`"
-            )
+            raise ImportError("`beautifulsoup4` package not found: `pip install beautifulsoup4`")
 
         super().__init__(*args, **kwargs)
         self.max_count = max_count
@@ -56,9 +50,9 @@ class MboxReader(BaseReader):
     def load_data(
         self,
         file: Path,
-        extra_info: Optional[Dict] = None,
-        fs: Optional[AbstractFileSystem] = None,
-    ) -> List[Document]:
+        extra_info: dict | None = None,
+        fs: AbstractFileSystem | None = None,
+    ) -> list[Document]:
         """Parse file into string."""
         # Import required libraries
         import mailbox
@@ -74,7 +68,7 @@ class MboxReader(BaseReader):
             )
 
         i = 0
-        results: List[str] = []
+        results: list[str] = []
         # Load file using mailbox
         bytes_parser = BytesParser(policy=default).parse
         mbox = mailbox.mbox(file, factory=bytes_parser)  # type: ignore
@@ -124,7 +118,7 @@ class MboxReader(BaseReader):
 class EmlxMboxReader(MboxReader):
     """
     EmlxMboxReader - Modified MboxReader that handles directories of .emlx files.
-    
+
     Extends MboxReader to work with Apple Mail's .emlx format by:
     1. Reading .emlx files from a directory
     2. Converting them to mbox format in memory
@@ -134,13 +128,13 @@ class EmlxMboxReader(MboxReader):
     def load_data(
         self,
         directory: Path,
-        extra_info: Optional[Dict] = None,
-        fs: Optional[AbstractFileSystem] = None,
-    ) -> List[Document]:
+        extra_info: dict | None = None,
+        fs: AbstractFileSystem | None = None,
+    ) -> list[Document]:
         """Parse .emlx files from directory into strings using MboxReader logic."""
-        import tempfile
         import os
-        
+        import tempfile
+
         if fs:
             logger.warning(
                 "fs was specified but EmlxMboxReader doesn't support loading "
@@ -150,37 +144,37 @@ class EmlxMboxReader(MboxReader):
         # Find all .emlx files in the directory
         emlx_files = list(directory.glob("*.emlx"))
         logger.info(f"Found {len(emlx_files)} .emlx files in {directory}")
-        
+
         if not emlx_files:
             logger.warning(f"No .emlx files found in {directory}")
             return []
 
         # Create a temporary mbox file
-        with tempfile.NamedTemporaryFile(mode='w', suffix='.mbox', delete=False) as temp_mbox:
+        with tempfile.NamedTemporaryFile(mode="w", suffix=".mbox", delete=False) as temp_mbox:
             temp_mbox_path = temp_mbox.name
-            
+
             # Convert .emlx files to mbox format
             for emlx_file in emlx_files:
                 try:
                     # Read the .emlx file
-                    with open(emlx_file, 'r', encoding='utf-8', errors='ignore') as f:
+                    with open(emlx_file, encoding="utf-8", errors="ignore") as f:
                         content = f.read()
-                    
+
                     # .emlx format: first line is length, rest is email content
-                    lines = content.split('\n', 1)
+                    lines = content.split("\n", 1)
                     if len(lines) >= 2:
                         email_content = lines[1]  # Skip the length line
-                        
+
                         # Write to mbox format (each message starts with "From " and ends with blank line)
                         temp_mbox.write(f"From {emlx_file.name} {email_content}\n\n")
-                    
+
                 except Exception as e:
                     logger.warning(f"Failed to process {emlx_file}: {e}")
                     continue
-            
+
             # Close the temporary file so MboxReader can read it
             temp_mbox.close()
-            
+
             try:
                 # Use the parent MboxReader's logic to parse the mbox file
                 return super().load_data(Path(temp_mbox_path), extra_info, fs)
@@ -188,5 +182,5 @@ class EmlxMboxReader(MboxReader):
                 # Clean up temporary file
                 try:
                     os.unlink(temp_mbox_path)
-                except:
-                    pass
+                except OSError:
+                    pass

apps/email_rag.py

@@ -0,0 +1,157 @@
+"""
+Email RAG example using the unified interface.
+Supports Apple Mail on macOS.
+"""
+
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+from base_rag_example import BaseRAGExample
+from chunking import create_text_chunks
+
+from .email_data.LEANN_email_reader import EmlxReader
+
+
+class EmailRAG(BaseRAGExample):
+    """RAG example for Apple Mail processing."""
+
+    def __init__(self):
+        # Set default values BEFORE calling super().__init__
+        self.max_items_default = -1  # Process all emails by default
+        self.embedding_model_default = (
+            "sentence-transformers/all-MiniLM-L6-v2"  # Fast 384-dim model
+        )
+
+        super().__init__(
+            name="Email",
+            description="Process and query Apple Mail emails with LEANN",
+            default_index_name="mail_index",
+        )
+
+    def _add_specific_arguments(self, parser):
+        """Add email-specific arguments."""
+        email_group = parser.add_argument_group("Email Parameters")
+        email_group.add_argument(
+            "--mail-path",
+            type=str,
+            default=None,
+            help="Path to Apple Mail directory (auto-detected if not specified)",
+        )
+        email_group.add_argument(
+            "--include-html", action="store_true", help="Include HTML content in email processing"
+        )
+        email_group.add_argument(
+            "--chunk-size", type=int, default=256, help="Text chunk size (default: 256)"
+        )
+        email_group.add_argument(
+            "--chunk-overlap", type=int, default=25, help="Text chunk overlap (default: 25)"
+        )
+
+    def _find_mail_directories(self) -> list[Path]:
+        """Auto-detect all Apple Mail directories."""
+        mail_base = Path.home() / "Library" / "Mail"
+        if not mail_base.exists():
+            return []
+
+        # Find all Messages directories
+        messages_dirs = []
+        for item in mail_base.rglob("Messages"):
+            if item.is_dir():
+                messages_dirs.append(item)
+
+        return messages_dirs
+
+    async def load_data(self, args) -> list[str]:
+        """Load emails and convert to text chunks."""
+        # Determine mail directories
+        if args.mail_path:
+            messages_dirs = [Path(args.mail_path)]
+        else:
+            print("Auto-detecting Apple Mail directories...")
+            messages_dirs = self._find_mail_directories()
+
+        if not messages_dirs:
+            print("No Apple Mail directories found!")
+            print("Please specify --mail-path manually")
+            return []
+
+        print(f"Found {len(messages_dirs)} mail directories")
+
+        # Create reader
+        reader = EmlxReader(include_html=args.include_html)
+
+        # Process each directory
+        all_documents = []
+        total_processed = 0
+
+        for i, messages_dir in enumerate(messages_dirs):
+            print(f"\nProcessing directory {i + 1}/{len(messages_dirs)}: {messages_dir}")
+
+            try:
+                # Count emlx files
+                emlx_files = list(messages_dir.glob("*.emlx"))
+                print(f"Found {len(emlx_files)} email files")
+
+                # Apply max_items limit per directory
+                max_per_dir = -1  # Default to process all
+                if args.max_items > 0:
+                    remaining = args.max_items - total_processed
+                    if remaining <= 0:
+                        break
+                    max_per_dir = remaining
+                # If args.max_items == -1, max_per_dir stays -1 (process all)
+
+                # Load emails - fix the parameter passing
+                documents = reader.load_data(
+                    input_dir=str(messages_dir),
+                    max_count=max_per_dir,
+                )
+
+                if documents:
+                    all_documents.extend(documents)
+                    total_processed += len(documents)
+                    print(f"Processed {len(documents)} emails from this directory")
+
+            except Exception as e:
+                print(f"Error processing {messages_dir}: {e}")
+                continue
+
+        if not all_documents:
+            print("No emails found to process!")
+            return []
+
+        print(f"\nTotal emails processed: {len(all_documents)}")
+        print("now starting to split into text chunks ... take some time")
+
+        # Convert to text chunks
+        # Email reader uses chunk_overlap=25 as in original
+        all_texts = create_text_chunks(
+            all_documents, chunk_size=args.chunk_size, chunk_overlap=args.chunk_overlap
+        )
+
+        return all_texts
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Check platform
+    if sys.platform != "darwin":
+        print("\n⚠️  Warning: This example is designed for macOS (Apple Mail)")
+        print("   Windows/Linux support coming soon!\n")
+
+    # Example queries for email RAG
+    print("\n📧 Email RAG Example")
+    print("=" * 50)
+    print("\nExample queries you can try:")
+    print("- 'What did my boss say about deadlines?'")
+    print("- 'Find emails about travel expenses'")
+    print("- 'Show me emails from last month about the project'")
+    print("- 'What food did I order from DoorDash?'")
+    print("\nNote: You may need to grant Full Disk Access to your terminal\n")
+
+    rag = EmailRAG()
+    asyncio.run(rag.run())

apps/history_data/__init__.py

@@ -1,3 +1,3 @@
 from .history import ChromeHistoryReader
 
-__all__ = ['ChromeHistoryReader'] 
+__all__ = ["ChromeHistoryReader"]

apps/history_data/history.py

@@ -1,77 +1,81 @@
-import sqlite3
 import os
+import sqlite3
 from pathlib import Path
-from typing import List, Any
+from typing import Any
+
 from llama_index.core import Document
 from llama_index.core.readers.base import BaseReader
 
+
 class ChromeHistoryReader(BaseReader):
     """
     Chrome browser history reader that extracts browsing data from SQLite database.
-    
+
     Reads Chrome history from the default Chrome profile location and creates documents
     with embedded metadata similar to the email reader structure.
     """
-    
+
     def __init__(self) -> None:
         """Initialize."""
         pass
-    
-    def load_data(self, input_dir: str = None, **load_kwargs: Any) -> List[Document]:
+
+    def load_data(self, input_dir: str | None = None, **load_kwargs: Any) -> list[Document]:
         """
         Load Chrome history data from the default Chrome profile location.
-        
+
         Args:
             input_dir: Not used for Chrome history (kept for compatibility)
             **load_kwargs:
                 max_count (int): Maximum amount of history entries to read.
                 chrome_profile_path (str): Custom path to Chrome profile directory.
         """
-        docs: List[Document] = []
-        max_count = load_kwargs.get('max_count', 1000)
-        chrome_profile_path = load_kwargs.get('chrome_profile_path', None)
-        
+        docs: list[Document] = []
+        max_count = load_kwargs.get("max_count", 1000)
+        chrome_profile_path = load_kwargs.get("chrome_profile_path", None)
+
         # Default Chrome profile path on macOS
         if chrome_profile_path is None:
-            chrome_profile_path = os.path.expanduser("~/Library/Application Support/Google/Chrome/Default")
-        
+            chrome_profile_path = os.path.expanduser(
+                "~/Library/Application Support/Google/Chrome/Default"
+            )
+
         history_db_path = os.path.join(chrome_profile_path, "History")
-        
+
         if not os.path.exists(history_db_path):
             print(f"Chrome history database not found at: {history_db_path}")
             return docs
-        
+
         try:
             # Connect to the Chrome history database
             print(f"Connecting to database: {history_db_path}")
             conn = sqlite3.connect(history_db_path)
             cursor = conn.cursor()
-            
+
             # Query to get browsing history with metadata (removed created_time column)
             query = """
-            SELECT 
+            SELECT
                 datetime(last_visit_time/1000000-11644473600,'unixepoch','localtime') as last_visit,
-                url, 
-                title, 
-                visit_count, 
-                typed_count, 
+                url,
+                title,
+                visit_count,
+                typed_count,
                 hidden
-            FROM urls 
+            FROM urls
             ORDER BY last_visit_time DESC
             """
-            
+
             print(f"Executing query on database: {history_db_path}")
             cursor.execute(query)
             rows = cursor.fetchall()
             print(f"Query returned {len(rows)} rows")
-            
+
             count = 0
             for row in rows:
                 if count >= max_count and max_count > 0:
                     break
-                
-                last_visit, url, title, visit_count, typed_count, hidden = row
-                
+
+                last_visit, url, title, visit_count, typed_count, _hidden = row
+
                 # Create document content with metadata embedded in text
                 doc_content = f"""
 [Title]: {title}
@@ -80,38 +84,43 @@ class ChromeHistoryReader(BaseReader):
 [Visit times]: {visit_count}
 [Typed times]: {typed_count}
 """
-                
+
                 # Create document with embedded metadata
-                doc = Document(text=doc_content, metadata={ "title": title[0:150]})
+                doc = Document(text=doc_content, metadata={"title": title[0:150]})
                 # if len(title) > 150:
                 #     print(f"Title is too long: {title}")
                 docs.append(doc)
                 count += 1
-            
+
             conn.close()
             print(f"Loaded {len(docs)} Chrome history documents")
-            
+
         except Exception as e:
             print(f"Error reading Chrome history: {e}")
+            # add you may need to close your browser to make the database file available
+            # also highlight in red
+            print(
+                "\033[91mYou may need to close your browser to make the database file available\033[0m"
+            )
             return docs
-        
+
         return docs
 
     @staticmethod
-    def find_chrome_profiles() -> List[Path]:
+    def find_chrome_profiles() -> list[Path]:
         """
         Find all Chrome profile directories.
-        
+
         Returns:
             List of Path objects pointing to Chrome profile directories
         """
         chrome_base_path = Path(os.path.expanduser("~/Library/Application Support/Google/Chrome"))
         profile_dirs = []
-        
+
         if not chrome_base_path.exists():
             print(f"Chrome directory not found at: {chrome_base_path}")
             return profile_dirs
-        
+
         # Find all profile directories
         for profile_dir in chrome_base_path.iterdir():
             if profile_dir.is_dir() and profile_dir.name != "System Profile":
@@ -119,53 +128,59 @@ class ChromeHistoryReader(BaseReader):
                 if history_path.exists():
                     profile_dirs.append(profile_dir)
                     print(f"Found Chrome profile: {profile_dir}")
-        
+
         print(f"Found {len(profile_dirs)} Chrome profiles")
         return profile_dirs
 
     @staticmethod
-    def export_history_to_file(output_file: str = "chrome_history_export.txt", max_count: int = 1000):
+    def export_history_to_file(
+        output_file: str = "chrome_history_export.txt", max_count: int = 1000
+    ):
         """
         Export Chrome history to a text file using the same SQL query format.
-        
+
         Args:
             output_file: Path to the output file
             max_count: Maximum number of entries to export
         """
-        chrome_profile_path = os.path.expanduser("~/Library/Application Support/Google/Chrome/Default")
+        chrome_profile_path = os.path.expanduser(
+            "~/Library/Application Support/Google/Chrome/Default"
+        )
         history_db_path = os.path.join(chrome_profile_path, "History")
-        
+
         if not os.path.exists(history_db_path):
             print(f"Chrome history database not found at: {history_db_path}")
             return
-        
+
         try:
             conn = sqlite3.connect(history_db_path)
             cursor = conn.cursor()
-            
+
             query = """
-            SELECT 
+            SELECT
                 datetime(last_visit_time/1000000-11644473600,'unixepoch','localtime') as last_visit,
-                url, 
-                title, 
-                visit_count, 
-                typed_count, 
+                url,
+                title,
+                visit_count,
+                typed_count,
                 hidden
-            FROM urls 
+            FROM urls
             ORDER BY last_visit_time DESC
             LIMIT ?
             """
-            
+
             cursor.execute(query, (max_count,))
             rows = cursor.fetchall()
-            
-            with open(output_file, 'w', encoding='utf-8') as f:
+
+            with open(output_file, "w", encoding="utf-8") as f:
                 for row in rows:
                     last_visit, url, title, visit_count, typed_count, hidden = row
-                    f.write(f"{last_visit}\t{url}\t{title}\t{visit_count}\t{typed_count}\t{hidden}\n")
-            
+                    f.write(
+                        f"{last_visit}\t{url}\t{title}\t{visit_count}\t{typed_count}\t{hidden}\n"
+                    )
+
             conn.close()
             print(f"Exported {len(rows)} history entries to {output_file}")
-            
+
         except Exception as e:
-            print(f"Error exporting Chrome history: {e}") 
+            print(f"Error exporting Chrome history: {e}")

apps/history_data/wechat_history.py

@@ -2,30 +2,31 @@ import json
 import os
 import re
 import subprocess
-import sys
 import time
+from datetime import datetime
 from pathlib import Path
-from typing import List, Any, Dict, Optional
+from typing import Any
+
 from llama_index.core import Document
 from llama_index.core.readers.base import BaseReader
-from datetime import datetime
+
 
 class WeChatHistoryReader(BaseReader):
     """
     WeChat chat history reader that extracts chat data from exported JSON files.
-    
+
     Reads WeChat chat history from exported JSON files (from wechat-exporter tool)
     and creates documents with embedded metadata similar to the Chrome history reader structure.
-    
+
     Also includes utilities for automatic WeChat chat history export.
     """
-    
+
     def __init__(self) -> None:
         """Initialize."""
         self.packages_dir = Path(__file__).parent.parent.parent / "packages"
         self.wechat_exporter_dir = self.packages_dir / "wechat-exporter"
         self.wechat_decipher_dir = self.packages_dir / "wechat-decipher-macos"
-    
+
     def check_wechat_running(self) -> bool:
         """Check if WeChat is currently running."""
         try:
@@ -33,24 +34,30 @@ class WeChatHistoryReader(BaseReader):
             return result.returncode == 0
         except Exception:
             return False
-    
+
     def install_wechattweak(self) -> bool:
         """Install WeChatTweak CLI tool."""
         try:
             # Create wechat-exporter directory if it doesn't exist
             self.wechat_exporter_dir.mkdir(parents=True, exist_ok=True)
-            
+
             wechattweak_path = self.wechat_exporter_dir / "wechattweak-cli"
             if not wechattweak_path.exists():
                 print("Downloading WeChatTweak CLI...")
-                subprocess.run([
-                    "curl", "-L", "-o", str(wechattweak_path),
-                    "https://github.com/JettChenT/WeChatTweak-CLI/releases/latest/download/wechattweak-cli"
-                ], check=True)
-            
+                subprocess.run(
+                    [
+                        "curl",
+                        "-L",
+                        "-o",
+                        str(wechattweak_path),
+                        "https://github.com/JettChenT/WeChatTweak-CLI/releases/latest/download/wechattweak-cli",
+                    ],
+                    check=True,
+                )
+
             # Make executable
             wechattweak_path.chmod(0o755)
-            
+
             # Install WeChatTweak
             print("Installing WeChatTweak...")
             subprocess.run(["sudo", str(wechattweak_path), "install"], check=True)
@@ -58,7 +65,7 @@ class WeChatHistoryReader(BaseReader):
         except Exception as e:
             print(f"Error installing WeChatTweak: {e}")
             return False
-    
+
     def restart_wechat(self):
         """Restart WeChat to apply WeChatTweak."""
         try:
@@ -69,302 +76,325 @@ class WeChatHistoryReader(BaseReader):
             time.sleep(5)  # Wait for WeChat to start
         except Exception as e:
             print(f"Error restarting WeChat: {e}")
-    
+
     def check_api_available(self) -> bool:
         """Check if WeChatTweak API is available."""
         try:
-            result = subprocess.run([
-                "curl", "-s", "http://localhost:48065/wechat/allcontacts"
-            ], capture_output=True, text=True, timeout=5)
+            result = subprocess.run(
+                ["curl", "-s", "http://localhost:48065/wechat/allcontacts"],
+                capture_output=True,
+                text=True,
+                timeout=5,
+            )
             return result.returncode == 0 and result.stdout.strip()
         except Exception:
             return False
-    
 
-
-    
     def _extract_readable_text(self, content: str) -> str:
         """
         Extract readable text from message content, removing XML and system messages.
-        
+
         Args:
             content: The raw message content (can be string or dict)
-            
+
         Returns:
             Cleaned, readable text
         """
         if not content:
             return ""
-        
+
         # Handle dictionary content (like quoted messages)
         if isinstance(content, dict):
             # Extract text from dictionary structure
             text_parts = []
-            if 'title' in content:
-                text_parts.append(str(content['title']))
-            if 'quoted' in content:
-                text_parts.append(str(content['quoted']))
-            if 'content' in content:
-                text_parts.append(str(content['content']))
-            if 'text' in content:
-                text_parts.append(str(content['text']))
-            
+            if "title" in content:
+                text_parts.append(str(content["title"]))
+            if "quoted" in content:
+                text_parts.append(str(content["quoted"]))
+            if "content" in content:
+                text_parts.append(str(content["content"]))
+            if "text" in content:
+                text_parts.append(str(content["text"]))
+
             if text_parts:
                 return " | ".join(text_parts)
             else:
                 # If we can't extract meaningful text from dict, return empty
                 return ""
-        
+
         # Handle string content
         if not isinstance(content, str):
             return ""
-        
+
         # Remove common prefixes like "wxid_xxx:\n"
-        clean_content = re.sub(r'^wxid_[^:]+:\s*', '', content)
-        clean_content = re.sub(r'^[^:]+:\s*', '', clean_content)
-        
+        clean_content = re.sub(r"^wxid_[^:]+:\s*", "", content)
+        clean_content = re.sub(r"^[^:]+:\s*", "", clean_content)
+
         # If it's just XML or system message, return empty
-        if clean_content.strip().startswith('<') or 'recalled a message' in clean_content:
+        if clean_content.strip().startswith("<") or "recalled a message" in clean_content:
             return ""
-        
+
         return clean_content.strip()
-    
+
     def _is_text_message(self, content: str) -> bool:
         """
         Check if a message contains readable text content.
-        
+
         Args:
             content: The message content (can be string or dict)
-            
+
         Returns:
             True if the message contains readable text, False otherwise
         """
         if not content:
             return False
-        
+
         # Handle dictionary content
         if isinstance(content, dict):
             # Check if dict has any readable text fields
-            text_fields = ['title', 'quoted', 'content', 'text']
+            text_fields = ["title", "quoted", "content", "text"]
             for field in text_fields:
-                if field in content and content[field]:
+                if content.get(field):
                     return True
             return False
-        
+
         # Handle string content
         if not isinstance(content, str):
             return False
-        
+
         # Skip image messages (contain XML with img tags)
-        if '<img' in content and 'cdnurl' in content:
+        if "<img" in content and "cdnurl" in content:
             return False
-        
+
         # Skip emoji messages (contain emoji XML tags)
-        if '<emoji' in content and 'productid' in content:
+        if "<emoji" in content and "productid" in content:
             return False
-        
+
         # Skip voice messages
-        if '<voice' in content:
+        if "<voice" in content:
             return False
-        
+
         # Skip video messages
-        if '<video' in content:
+        if "<video" in content:
             return False
-        
+
         # Skip file messages
-        if '<appmsg' in content and 'appid' in content:
+        if "<appmsg" in content and "appid" in content:
             return False
-        
+
         # Skip system messages (like "recalled a message")
-        if 'recalled a message' in content:
+        if "recalled a message" in content:
             return False
-        
+
         # Check if there's actual readable text (not just XML or system messages)
         # Remove common prefixes like "wxid_xxx:\n" and check for actual content
-        clean_content = re.sub(r'^wxid_[^:]+:\s*', '', content)
-        clean_content = re.sub(r'^[^:]+:\s*', '', clean_content)
-        
+        clean_content = re.sub(r"^wxid_[^:]+:\s*", "", content)
+        clean_content = re.sub(r"^[^:]+:\s*", "", clean_content)
+
         # If after cleaning we have meaningful text, consider it readable
-        if len(clean_content.strip()) > 0 and not clean_content.strip().startswith('<'):
+        if len(clean_content.strip()) > 0 and not clean_content.strip().startswith("<"):
             return True
-        
+
         return False
-    
-    def _concatenate_messages(self, messages: List[Dict], max_length: int = 128, 
-                             time_window_minutes: int = 30, overlap_messages: int = 0) -> List[Dict]:
+
+    def _concatenate_messages(
+        self,
+        messages: list[dict],
+        max_length: int = 128,
+        time_window_minutes: int = 30,
+        overlap_messages: int = 0,
+    ) -> list[dict]:
         """
         Concatenate messages based on length and time rules.
-        
+
         Args:
             messages: List of message dictionaries
             max_length: Maximum length for concatenated message groups. Use -1 to disable length constraint.
             time_window_minutes: Time window in minutes to group messages together. Use -1 to disable time constraint.
             overlap_messages: Number of messages to overlap between consecutive groups
-            
+
         Returns:
             List of concatenated message groups
         """
         if not messages:
             return []
-        
+
         concatenated_groups = []
         current_group = []
         current_length = 0
         last_timestamp = None
-        
+
         for message in messages:
             # Extract message info
-            content = message.get('content', '')
-            message_text = message.get('message', '')
-            create_time = message.get('createTime', 0)
-            from_user = message.get('fromUser', '')
-            to_user = message.get('toUser', '')
-            is_sent_from_self = message.get('isSentFromSelf', False)
-            
+            content = message.get("content", "")
+            message_text = message.get("message", "")
+            create_time = message.get("createTime", 0)
+            message.get("fromUser", "")
+            message.get("toUser", "")
+            message.get("isSentFromSelf", False)
+
             # Extract readable text
             readable_text = self._extract_readable_text(content)
             if not readable_text:
                 readable_text = message_text
-            
+
             # Skip empty messages
             if not readable_text.strip():
                 continue
-            
+
             # Check time window constraint (only if time_window_minutes != -1)
             if time_window_minutes != -1 and last_timestamp is not None and create_time > 0:
                 time_diff_minutes = (create_time - last_timestamp) / 60
                 if time_diff_minutes > time_window_minutes:
                     # Time gap too large, start new group
                     if current_group:
-                        concatenated_groups.append({
-                            'messages': current_group,
-                            'total_length': current_length,
-                            'start_time': current_group[0].get('createTime', 0),
-                            'end_time': current_group[-1].get('createTime', 0)
-                        })
+                        concatenated_groups.append(
+                            {
+                                "messages": current_group,
+                                "total_length": current_length,
+                                "start_time": current_group[0].get("createTime", 0),
+                                "end_time": current_group[-1].get("createTime", 0),
+                            }
+                        )
                         # Keep last few messages for overlap
                         if overlap_messages > 0 and len(current_group) > overlap_messages:
                             current_group = current_group[-overlap_messages:]
-                            current_length = sum(len(self._extract_readable_text(msg.get('content', '')) or msg.get('message', '')) for msg in current_group)
+                            current_length = sum(
+                                len(
+                                    self._extract_readable_text(msg.get("content", ""))
+                                    or msg.get("message", "")
+                                )
+                                for msg in current_group
+                            )
                         else:
                             current_group = []
                             current_length = 0
-            
+
             # Check length constraint (only if max_length != -1)
             message_length = len(readable_text)
             if max_length != -1 and current_length + message_length > max_length and current_group:
                 # Current group would exceed max length, save it and start new
-                concatenated_groups.append({
-                    'messages': current_group,
-                    'total_length': current_length,
-                    'start_time': current_group[0].get('createTime', 0),
-                    'end_time': current_group[-1].get('createTime', 0)
-                })
+                concatenated_groups.append(
+                    {
+                        "messages": current_group,
+                        "total_length": current_length,
+                        "start_time": current_group[0].get("createTime", 0),
+                        "end_time": current_group[-1].get("createTime", 0),
+                    }
+                )
                 # Keep last few messages for overlap
                 if overlap_messages > 0 and len(current_group) > overlap_messages:
                     current_group = current_group[-overlap_messages:]
-                    current_length = sum(len(self._extract_readable_text(msg.get('content', '')) or msg.get('message', '')) for msg in current_group)
+                    current_length = sum(
+                        len(
+                            self._extract_readable_text(msg.get("content", ""))
+                            or msg.get("message", "")
+                        )
+                        for msg in current_group
+                    )
                 else:
                     current_group = []
                     current_length = 0
-            
+
             # Add message to current group
             current_group.append(message)
             current_length += message_length
             last_timestamp = create_time
-        
+
         # Add the last group if it exists
         if current_group:
-            concatenated_groups.append({
-                'messages': current_group,
-                'total_length': current_length,
-                'start_time': current_group[0].get('createTime', 0),
-                'end_time': current_group[-1].get('createTime', 0)
-            })
-        
+            concatenated_groups.append(
+                {
+                    "messages": current_group,
+                    "total_length": current_length,
+                    "start_time": current_group[0].get("createTime", 0),
+                    "end_time": current_group[-1].get("createTime", 0),
+                }
+            )
+
         return concatenated_groups
-    
-    def _create_concatenated_content(self, message_group: Dict, contact_name: str) -> str:
+
+    def _create_concatenated_content(self, message_group: dict, contact_name: str) -> str:
         """
         Create concatenated content from a group of messages.
-        
+
         Args:
             message_group: Dictionary containing messages and metadata
             contact_name: Name of the contact
-            
+
         Returns:
             Formatted concatenated content
         """
-        messages = message_group['messages']
-        start_time = message_group['start_time']
-        end_time = message_group['end_time']
-        
+        messages = message_group["messages"]
+        start_time = message_group["start_time"]
+        end_time = message_group["end_time"]
+
         # Format timestamps
         if start_time:
             try:
                 start_timestamp = datetime.fromtimestamp(start_time)
-                start_time_str = start_timestamp.strftime('%Y-%m-%d %H:%M:%S')
-            except:
+                start_time_str = start_timestamp.strftime("%Y-%m-%d %H:%M:%S")
+            except (ValueError, OSError):
                 start_time_str = str(start_time)
         else:
             start_time_str = "Unknown"
-        
+
         if end_time:
             try:
                 end_timestamp = datetime.fromtimestamp(end_time)
-                end_time_str = end_timestamp.strftime('%Y-%m-%d %H:%M:%S')
-            except:
+                end_time_str = end_timestamp.strftime("%Y-%m-%d %H:%M:%S")
+            except (ValueError, OSError):
                 end_time_str = str(end_time)
         else:
             end_time_str = "Unknown"
-        
+
         # Build concatenated message content
         message_parts = []
         for message in messages:
-            content = message.get('content', '')
-            message_text = message.get('message', '')
-            create_time = message.get('createTime', 0)
-            is_sent_from_self = message.get('isSentFromSelf', False)
-            
+            content = message.get("content", "")
+            message_text = message.get("message", "")
+            create_time = message.get("createTime", 0)
+            is_sent_from_self = message.get("isSentFromSelf", False)
+
             # Extract readable text
             readable_text = self._extract_readable_text(content)
             if not readable_text:
                 readable_text = message_text
-            
+
             # Format individual message
             if create_time:
                 try:
                     timestamp = datetime.fromtimestamp(create_time)
                     # change to YYYY-MM-DD HH:MM:SS
-                    time_str = timestamp.strftime('%Y-%m-%d %H:%M:%S')
-                except:
+                    time_str = timestamp.strftime("%Y-%m-%d %H:%M:%S")
+                except (ValueError, OSError):
                     time_str = str(create_time)
             else:
                 time_str = "Unknown"
-            
+
             sender = "[Me]" if is_sent_from_self else "[Contact]"
             message_parts.append(f"({time_str}) {sender}: {readable_text}")
-        
+
         concatenated_text = "\n".join(message_parts)
-        
+
         # Create final document content
         doc_content = f"""
 Contact: {contact_name}
 Time Range: {start_time_str} - {end_time_str}
-Messages ({len(messages)} messages, {message_group['total_length']} chars):
+Messages ({len(messages)} messages, {message_group["total_length"]} chars):
 
 {concatenated_text}
 """
-        # TODO @yichuan give better format and rich info here!    
+        # TODO @yichuan give better format and rich info here!
         doc_content = f"""
 {concatenated_text}
 """
         return doc_content, contact_name
-    
-    def load_data(self, input_dir: str = None, **load_kwargs: Any) -> List[Document]:
+
+    def load_data(self, input_dir: str | None = None, **load_kwargs: Any) -> list[Document]:
         """
         Load WeChat chat history data from exported JSON files.
-        
+
         Args:
             input_dir: Directory containing exported WeChat JSON files
             **load_kwargs:
@@ -376,97 +406,104 @@ Messages ({len(messages)} messages, {message_group['total_length']} chars):
                 time_window_minutes (int): Time window in minutes to group messages together (default: 30).
                 overlap_messages (int): Number of messages to overlap between consecutive groups (default: 2).
         """
-        docs: List[Document] = []
-        max_count = load_kwargs.get('max_count', 1000)
-        wechat_export_dir = load_kwargs.get('wechat_export_dir', None)
-        include_non_text = load_kwargs.get('include_non_text', False)
-        concatenate_messages = load_kwargs.get('concatenate_messages', False)
-        max_length = load_kwargs.get('max_length', 1000)
-        time_window_minutes = load_kwargs.get('time_window_minutes', 30)
-        
+        docs: list[Document] = []
+        max_count = load_kwargs.get("max_count", 1000)
+        wechat_export_dir = load_kwargs.get("wechat_export_dir", None)
+        include_non_text = load_kwargs.get("include_non_text", False)
+        concatenate_messages = load_kwargs.get("concatenate_messages", False)
+        max_length = load_kwargs.get("max_length", 1000)
+        time_window_minutes = load_kwargs.get("time_window_minutes", 30)
+
         # Default WeChat export path
         if wechat_export_dir is None:
             wechat_export_dir = "./wechat_export_test"
-        
+
         if not os.path.exists(wechat_export_dir):
             print(f"WeChat export directory not found at: {wechat_export_dir}")
             return docs
-        
+
         try:
             # Find all JSON files in the export directory
             json_files = list(Path(wechat_export_dir).glob("*.json"))
             print(f"Found {len(json_files)} WeChat chat history files")
-            
+
             count = 0
             for json_file in json_files:
                 if count >= max_count and max_count > 0:
                     break
-                
+
                 try:
-                    with open(json_file, 'r', encoding='utf-8') as f:
+                    with open(json_file, encoding="utf-8") as f:
                         chat_data = json.load(f)
-                    
+
                     # Extract contact name from filename
                     contact_name = json_file.stem
-                    
+
                     if concatenate_messages:
                         # Filter messages to only include readable text messages
                         readable_messages = []
                         for message in chat_data:
                             try:
-                                content = message.get('content', '')
+                                content = message.get("content", "")
                                 if not include_non_text and not self._is_text_message(content):
                                     continue
-                                
+
                                 readable_text = self._extract_readable_text(content)
                                 if not readable_text and not include_non_text:
                                     continue
-                                
+
                                 readable_messages.append(message)
                             except Exception as e:
                                 print(f"Error processing message in {json_file}: {e}")
                                 continue
-                        
+
                         # Concatenate messages based on rules
                         message_groups = self._concatenate_messages(
-                            readable_messages, 
-                            max_length=-1, 
-                            time_window_minutes=-1,
-                            overlap_messages=0  # Keep 2 messages overlap between groups
+                            readable_messages,
+                            max_length=max_length,
+                            time_window_minutes=time_window_minutes,
+                            overlap_messages=0,  # No overlap between groups
                         )
-                        
+
                         # Create documents from concatenated groups
                         for message_group in message_groups:
                             if count >= max_count and max_count > 0:
                                 break
-                            
-                            doc_content, contact_name  = self._create_concatenated_content(message_group, contact_name)
-                            doc = Document(text=doc_content, metadata={"contact_name": contact_name})
+
+                            doc_content, contact_name = self._create_concatenated_content(
+                                message_group, contact_name
+                            )
+                            doc = Document(
+                                text=doc_content,
+                                metadata={"contact_name": contact_name},
+                            )
                             docs.append(doc)
                             count += 1
-                        
-                        print(f"Created {len(message_groups)} concatenated message groups for {contact_name}")
-                        
+
+                        print(
+                            f"Created {len(message_groups)} concatenated message groups for {contact_name}"
+                        )
+
                     else:
                         # Original single-message processing
                         for message in chat_data:
                             if count >= max_count and max_count > 0:
                                 break
-                            
+
                             # Extract message information
-                            from_user = message.get('fromUser', '')
-                            to_user = message.get('toUser', '')
-                            content = message.get('content', '')
-                            message_text = message.get('message', '')
-                            create_time = message.get('createTime', 0)
-                            is_sent_from_self = message.get('isSentFromSelf', False)
-                            
+                            message.get("fromUser", "")
+                            message.get("toUser", "")
+                            content = message.get("content", "")
+                            message_text = message.get("message", "")
+                            create_time = message.get("createTime", 0)
+                            is_sent_from_self = message.get("isSentFromSelf", False)
+
                             # Handle content that might be dict or string
                             try:
                                 # Check if this is a readable text message
                                 if not include_non_text and not self._is_text_message(content):
                                     continue
-                                
+
                                 # Extract readable text
                                 readable_text = self._extract_readable_text(content)
                                 if not readable_text and not include_non_text:
@@ -475,17 +512,17 @@ Messages ({len(messages)} messages, {message_group['total_length']} chars):
                                 # Skip messages that cause processing errors
                                 print(f"Error processing message in {json_file}: {e}")
                                 continue
-                            
+
                             # Convert timestamp to readable format
                             if create_time:
                                 try:
                                     timestamp = datetime.fromtimestamp(create_time)
-                                    time_str = timestamp.strftime('%Y-%m-%d %H:%M:%S')
-                                except:
+                                    time_str = timestamp.strftime("%Y-%m-%d %H:%M:%S")
+                                except (ValueError, OSError):
                                     time_str = str(create_time)
                             else:
                                 time_str = "Unknown"
-                            
+
                             # Create document content with metadata header and contact info
                             doc_content = f"""
 Contact: {contact_name}
@@ -493,57 +530,66 @@ Is sent from self: {is_sent_from_self}
 Time: {time_str}
 Message: {readable_text if readable_text else message_text}
 """
-                            
+
                             # Create document with embedded metadata
-                            doc = Document(text=doc_content, metadata={})
+                            doc = Document(
+                                text=doc_content, metadata={"contact_name": contact_name}
+                            )
                             docs.append(doc)
                             count += 1
-                        
+
                 except Exception as e:
                     print(f"Error reading {json_file}: {e}")
                     continue
-            
+
             print(f"Loaded {len(docs)} WeChat chat documents")
-            
+
         except Exception as e:
             print(f"Error reading WeChat history: {e}")
             return docs
-        
+
         return docs
 
     @staticmethod
-    def find_wechat_export_dirs() -> List[Path]:
+    def find_wechat_export_dirs() -> list[Path]:
         """
         Find all WeChat export directories.
-        
+
         Returns:
             List of Path objects pointing to WeChat export directories
         """
         export_dirs = []
-        
+
         # Look for common export directory names
         possible_dirs = [
-            Path("./wechat_export_test"),
             Path("./wechat_export"),
+            Path("./wechat_export_direct"),
             Path("./wechat_chat_history"),
-            Path("./chat_export")
+            Path("./chat_export"),
         ]
-        
+
         for export_dir in possible_dirs:
             if export_dir.exists() and export_dir.is_dir():
                 json_files = list(export_dir.glob("*.json"))
                 if json_files:
                     export_dirs.append(export_dir)
-                    print(f"Found WeChat export directory: {export_dir} with {len(json_files)} files")
-        
+                    print(
+                        f"Found WeChat export directory: {export_dir} with {len(json_files)} files"
+                    )
+
         print(f"Found {len(export_dirs)} WeChat export directories")
         return export_dirs
 
     @staticmethod
-    def export_chat_to_file(output_file: str = "wechat_chat_export.txt", max_count: int = 1000, export_dir: str = None, include_non_text: bool = False):
+    def export_chat_to_file(
+        output_file: str = "wechat_chat_export.txt",
+        max_count: int = 1000,
+        export_dir: str | None = None,
+        include_non_text: bool = False,
+    ):
         """
         Export WeChat chat history to a text file.
-        
+
         Args:
             output_file: Path to the output file
             max_count: Maximum number of entries to export
@@ -552,36 +598,36 @@ Message: {readable_text if readable_text else message_text}
         """
         if export_dir is None:
             export_dir = "./wechat_export_test"
-        
+
         if not os.path.exists(export_dir):
             print(f"WeChat export directory not found at: {export_dir}")
             return
-        
+
         try:
             json_files = list(Path(export_dir).glob("*.json"))
-            
-            with open(output_file, 'w', encoding='utf-8') as f:
+
+            with open(output_file, "w", encoding="utf-8") as f:
                 count = 0
                 for json_file in json_files:
                     if count >= max_count and max_count > 0:
                         break
-                    
+
                     try:
-                        with open(json_file, 'r', encoding='utf-8') as json_f:
+                        with open(json_file, encoding="utf-8") as json_f:
                             chat_data = json.load(json_f)
-                        
+
                         contact_name = json_file.stem
                         f.write(f"\n=== Chat with {contact_name} ===\n")
-                        
+
                         for message in chat_data:
                             if count >= max_count and max_count > 0:
                                 break
-                            
-                            from_user = message.get('fromUser', '')
-                            content = message.get('content', '')
-                            message_text = message.get('message', '')
-                            create_time = message.get('createTime', 0)
-                            
+
+                            from_user = message.get("fromUser", "")
+                            content = message.get("content", "")
+                            message_text = message.get("message", "")
+                            create_time = message.get("createTime", 0)
+
                             # Skip non-text messages unless requested
                             if not include_non_text:
                                 reader = WeChatHistoryReader()
@@ -591,83 +637,90 @@ Message: {readable_text if readable_text else message_text}
                                 if not readable_text:
                                     continue
                                 message_text = readable_text
-                            
+
                             if create_time:
                                 try:
                                     timestamp = datetime.fromtimestamp(create_time)
-                                    time_str = timestamp.strftime('%Y-%m-%d %H:%M:%S')
-                                except:
+                                    time_str = timestamp.strftime("%Y-%m-%d %H:%M:%S")
+                                except (ValueError, OSError):
                                     time_str = str(create_time)
                             else:
                                 time_str = "Unknown"
-                            
+
                             f.write(f"[{time_str}] {from_user}: {message_text}\n")
                             count += 1
-                            
+
                     except Exception as e:
                         print(f"Error processing {json_file}: {e}")
                         continue
-            
+
             print(f"Exported {count} chat entries to {output_file}")
-            
+
         except Exception as e:
             print(f"Error exporting WeChat chat history: {e}")
 
-    def export_wechat_chat_history(self, export_dir: str = "./wechat_export_direct") -> Optional[Path]:
+    def export_wechat_chat_history(self, export_dir: str = "./wechat_export_direct") -> Path | None:
         """
         Export WeChat chat history using wechat-exporter tool.
-        
+
         Args:
             export_dir: Directory to save exported chat history
-            
+
         Returns:
             Path to export directory if successful, None otherwise
         """
         try:
             import subprocess
             import sys
-            
+
             # Create export directory
             export_path = Path(export_dir)
             export_path.mkdir(exist_ok=True)
-            
+
             print(f"Exporting WeChat chat history to {export_path}...")
-            
+
             # Check if wechat-exporter directory exists
             if not self.wechat_exporter_dir.exists():
                 print(f"wechat-exporter directory not found at: {self.wechat_exporter_dir}")
                 return None
-            
+
             # Install requirements if needed
             requirements_file = self.wechat_exporter_dir / "requirements.txt"
             if requirements_file.exists():
                 print("Installing wechat-exporter requirements...")
-                subprocess.run([
-                    "uv", "pip", "install", "-r", str(requirements_file)
-                ], check=True)
-            
+                subprocess.run(["uv", "pip", "install", "-r", str(requirements_file)], check=True)
+
             # Run the export command
             print("Running wechat-exporter...")
-            result = subprocess.run([
-                sys.executable, str(self.wechat_exporter_dir / "main.py"), 
-                "export-all", str(export_path)
-            ], capture_output=True, text=True, check=True)
-            
+            result = subprocess.run(
+                [
+                    sys.executable,
+                    str(self.wechat_exporter_dir / "main.py"),
+                    "export-all",
+                    str(export_path),
+                ],
+                capture_output=True,
+                text=True,
+                check=True,
+            )
+
             print("Export command output:")
             print(result.stdout)
             if result.stderr:
                 print("Export errors:")
                 print(result.stderr)
-            
+
             # Check if export was successful
             if export_path.exists() and any(export_path.glob("*.json")):
                 json_files = list(export_path.glob("*.json"))
-                print(f"Successfully exported {len(json_files)} chat history files to {export_path}")
+                print(
+                    f"Successfully exported {len(json_files)} chat history files to {export_path}"
+                )
                 return export_path
             else:
                 print("Export completed but no JSON files found")
                 return None
-                
+
         except subprocess.CalledProcessError as e:
             print(f"Export command failed: {e}")
             print(f"Command output: {e.stdout}")
@@ -678,18 +731,18 @@ Message: {readable_text if readable_text else message_text}
             print("Please ensure WeChat is running and WeChatTweak is installed.")
             return None
 
-    def find_or_export_wechat_data(self, export_dir: str = "./wechat_export_direct") -> List[Path]:
+    def find_or_export_wechat_data(self, export_dir: str = "./wechat_export_direct") -> list[Path]:
         """
         Find existing WeChat exports or create new ones.
-        
+
         Args:
             export_dir: Directory to save exported chat history if needed
-            
+
         Returns:
             List of Path objects pointing to WeChat export directories
         """
         export_dirs = []
-        
+
         # Look for existing exports in common locations
         possible_export_dirs = [
             Path("./wechat_database_export"),
@@ -697,23 +750,25 @@ Message: {readable_text if readable_text else message_text}
             Path("./wechat_export"),
             Path("./wechat_export_direct"),
             Path("./wechat_chat_history"),
-            Path("./chat_export")
+            Path("./chat_export"),
         ]
-        
+
         for export_dir_path in possible_export_dirs:
             if export_dir_path.exists() and any(export_dir_path.glob("*.json")):
                 export_dirs.append(export_dir_path)
                 print(f"Found existing export: {export_dir_path}")
-        
+
         # If no existing exports, try to export automatically
         if not export_dirs:
             print("No existing WeChat exports found. Starting direct export...")
-            
+
             # Try to export using wechat-exporter
             exported_path = self.export_wechat_chat_history(export_dir)
             if exported_path:
                 export_dirs = [exported_path]
             else:
-                print("Failed to export WeChat data. Please ensure WeChat is running and WeChatTweak is installed.")
-        
-        return export_dirs 
+                print(
+                    "Failed to export WeChat data. Please ensure WeChat is running and WeChatTweak is installed."
+                )
+
+        return export_dirs

apps/imessage_data/__init__.py

@@ -0,0 +1 @@
+"""iMessage data processing module."""

apps/imessage_data/imessage_reader.py

@@ -0,0 +1,342 @@
+"""
+iMessage data reader.
+
+Reads and processes iMessage conversation data from the macOS Messages database.
+"""
+
+import sqlite3
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+from llama_index.core import Document
+from llama_index.core.readers.base import BaseReader
+
+
+class IMessageReader(BaseReader):
+    """
+    iMessage data reader.
+
+    Reads iMessage conversation data from the macOS Messages database (chat.db).
+    Processes conversations into structured documents with metadata.
+    """
+
+    def __init__(self, concatenate_conversations: bool = True) -> None:
+        """
+        Initialize.
+
+        Args:
+            concatenate_conversations: Whether to concatenate messages within conversations for better context
+        """
+        self.concatenate_conversations = concatenate_conversations
+
+    def _get_default_chat_db_path(self) -> Path:
+        """
+        Get the default path to the iMessage chat database.
+
+        Returns:
+            Path to the chat.db file
+        """
+        home = Path.home()
+        return home / "Library" / "Messages" / "chat.db"
+
+    def _convert_cocoa_timestamp(self, cocoa_timestamp: int) -> str:
+        """
+        Convert Cocoa timestamp to readable format.
+
+        Args:
+            cocoa_timestamp: Timestamp in Cocoa format (nanoseconds since 2001-01-01)
+
+        Returns:
+            Formatted timestamp string
+        """
+        if cocoa_timestamp == 0:
+            return "Unknown"
+
+        try:
+            # Cocoa timestamp is nanoseconds since 2001-01-01 00:00:00 UTC
+            # Convert to seconds and add to Unix epoch
+            cocoa_epoch = datetime(2001, 1, 1)
+            unix_timestamp = cocoa_timestamp / 1_000_000_000  # Convert nanoseconds to seconds
+            message_time = cocoa_epoch.timestamp() + unix_timestamp
+            return datetime.fromtimestamp(message_time).strftime("%Y-%m-%d %H:%M:%S")
+        except (ValueError, OSError):
+            return "Unknown"
+
+    def _get_contact_name(self, handle_id: str) -> str:
+        """
+        Get a readable contact name from handle ID.
+
+        Args:
+            handle_id: The handle ID (phone number or email)
+
+        Returns:
+            Formatted contact name
+        """
+        if not handle_id:
+            return "Unknown"
+
+        # Clean up phone numbers and emails for display
+        if "@" in handle_id:
+            return handle_id  # Email address
+        elif handle_id.startswith("+"):
+            return handle_id  # International phone number
+        else:
+            # Try to format as phone number
+            digits = "".join(filter(str.isdigit, handle_id))
+            if len(digits) == 10:
+                return f"({digits[:3]}) {digits[3:6]}-{digits[6:]}"
+            elif len(digits) == 11 and digits[0] == "1":
+                return f"+1 ({digits[1:4]}) {digits[4:7]}-{digits[7:]}"
+            else:
+                return handle_id
+
+    def _read_messages_from_db(self, db_path: Path) -> list[dict]:
+        """
+        Read messages from the iMessage database.
+
+        Args:
+            db_path: Path to the chat.db file
+
+        Returns:
+            List of message dictionaries
+        """
+        if not db_path.exists():
+            print(f"iMessage database not found at: {db_path}")
+            return []
+
+        try:
+            # Connect to the database
+            conn = sqlite3.connect(str(db_path))
+            cursor = conn.cursor()
+
+            # Query to get messages with chat and handle information
+            query = """
+            SELECT
+                m.ROWID as message_id,
+                m.text,
+                m.date,
+                m.is_from_me,
+                m.service,
+                c.chat_identifier,
+                c.display_name as chat_display_name,
+                h.id as handle_id,
+                c.ROWID as chat_id
+            FROM message m
+            LEFT JOIN chat_message_join cmj ON m.ROWID = cmj.message_id
+            LEFT JOIN chat c ON cmj.chat_id = c.ROWID
+            LEFT JOIN handle h ON m.handle_id = h.ROWID
+            WHERE m.text IS NOT NULL AND m.text != ''
+            ORDER BY c.ROWID, m.date
+            """
+
+            cursor.execute(query)
+            rows = cursor.fetchall()
+
+            messages = []
+            for row in rows:
+                (
+                    message_id,
+                    text,
+                    date,
+                    is_from_me,
+                    service,
+                    chat_identifier,
+                    chat_display_name,
+                    handle_id,
+                    chat_id,
+                ) = row
+
+                message = {
+                    "message_id": message_id,
+                    "text": text,
+                    "timestamp": self._convert_cocoa_timestamp(date),
+                    "is_from_me": bool(is_from_me),
+                    "service": service or "iMessage",
+                    "chat_identifier": chat_identifier or "Unknown",
+                    "chat_display_name": chat_display_name or "Unknown Chat",
+                    "handle_id": handle_id or "Unknown",
+                    "contact_name": self._get_contact_name(handle_id or ""),
+                    "chat_id": chat_id,
+                }
+                messages.append(message)
+
+            conn.close()
+            print(f"Found {len(messages)} messages in database")
+            return messages
+
+        except sqlite3.Error as e:
+            print(f"Error reading iMessage database: {e}")
+            return []
+        except Exception as e:
+            print(f"Unexpected error reading iMessage database: {e}")
+            return []
+
+    def _group_messages_by_chat(self, messages: list[dict]) -> dict[int, list[dict]]:
+        """
+        Group messages by chat ID.
+
+        Args:
+            messages: List of message dictionaries
+
+        Returns:
+            Dictionary mapping chat_id to list of messages
+        """
+        chats = {}
+        for message in messages:
+            chat_id = message["chat_id"]
+            if chat_id not in chats:
+                chats[chat_id] = []
+            chats[chat_id].append(message)
+
+        return chats
+
+    def _create_concatenated_content(self, chat_id: int, messages: list[dict]) -> str:
+        """
+        Create concatenated content from chat messages.
+
+        Args:
+            chat_id: The chat ID
+            messages: List of messages in the chat
+
+        Returns:
+            Concatenated text content
+        """
+        if not messages:
+            return ""
+
+        # Get chat info from first message
+        first_msg = messages[0]
+        chat_name = first_msg["chat_display_name"]
+        chat_identifier = first_msg["chat_identifier"]
+
+        # Build message content
+        message_parts = []
+        for message in messages:
+            timestamp = message["timestamp"]
+            is_from_me = message["is_from_me"]
+            text = message["text"]
+            contact_name = message["contact_name"]
+
+            if is_from_me:
+                prefix = "[You]"
+            else:
+                prefix = f"[{contact_name}]"
+
+            if timestamp != "Unknown":
+                prefix += f" ({timestamp})"
+
+            message_parts.append(f"{prefix}: {text}")
+
+        concatenated_text = "\n\n".join(message_parts)
+
+        doc_content = f"""Chat: {chat_name}
+Identifier: {chat_identifier}
+Messages ({len(messages)} messages):
+
+{concatenated_text}
+"""
+        return doc_content
+
+    def _create_individual_content(self, message: dict) -> str:
+        """
+        Create content for individual message.
+
+        Args:
+            message: Message dictionary
+
+        Returns:
+            Formatted message content
+        """
+        timestamp = message["timestamp"]
+        is_from_me = message["is_from_me"]
+        text = message["text"]
+        contact_name = message["contact_name"]
+        chat_name = message["chat_display_name"]
+
+        sender = "You" if is_from_me else contact_name
+
+        return f"""Message from {sender} in chat "{chat_name}"
+Time: {timestamp}
+Content: {text}
+"""
+
+    def load_data(self, input_dir: str | None = None, **load_kwargs: Any) -> list[Document]:
+        """
+        Load iMessage data and return as documents.
+
+        Args:
+            input_dir: Optional path to directory containing chat.db file.
+                      If not provided, uses default macOS location.
+            **load_kwargs: Additional arguments (unused)
+
+        Returns:
+            List of Document objects containing iMessage data
+        """
+        docs = []
+
+        # Determine database path
+        if input_dir:
+            db_path = Path(input_dir) / "chat.db"
+        else:
+            db_path = self._get_default_chat_db_path()
+
+        print(f"Reading iMessage database from: {db_path}")
+
+        # Read messages from database
+        messages = self._read_messages_from_db(db_path)
+        if not messages:
+            return docs
+
+        if self.concatenate_conversations:
+            # Group messages by chat and create concatenated documents
+            chats = self._group_messages_by_chat(messages)
+
+            for chat_id, chat_messages in chats.items():
+                if not chat_messages:
+                    continue
+
+                content = self._create_concatenated_content(chat_id, chat_messages)
+
+                # Create metadata
+                first_msg = chat_messages[0]
+                last_msg = chat_messages[-1]
+
+                metadata = {
+                    "source": "iMessage",
+                    "chat_id": chat_id,
+                    "chat_name": first_msg["chat_display_name"],
+                    "chat_identifier": first_msg["chat_identifier"],
+                    "message_count": len(chat_messages),
+                    "first_message_date": first_msg["timestamp"],
+                    "last_message_date": last_msg["timestamp"],
+                    "participants": list(
+                        {msg["contact_name"] for msg in chat_messages if not msg["is_from_me"]}
+                    ),
+                }
+
+                doc = Document(text=content, metadata=metadata)
+                docs.append(doc)
+
+        else:
+            # Create individual documents for each message
+            for message in messages:
+                content = self._create_individual_content(message)
+
+                metadata = {
+                    "source": "iMessage",
+                    "message_id": message["message_id"],
+                    "chat_id": message["chat_id"],
+                    "chat_name": message["chat_display_name"],
+                    "chat_identifier": message["chat_identifier"],
+                    "timestamp": message["timestamp"],
+                    "is_from_me": message["is_from_me"],
+                    "contact_name": message["contact_name"],
+                    "service": message["service"],
+                }
+
+                doc = Document(text=content, metadata=metadata)
+                docs.append(doc)
+
+        print(f"Created {len(docs)} documents from iMessage data")
+        return docs

apps/imessage_rag.py

@@ -0,0 +1,125 @@
+"""
+iMessage RAG Example.
+
+This example demonstrates how to build a RAG system on your iMessage conversation history.
+"""
+
+import asyncio
+from pathlib import Path
+
+from leann.chunking_utils import create_text_chunks
+
+from apps.base_rag_example import BaseRAGExample
+from apps.imessage_data.imessage_reader import IMessageReader
+
+
+class IMessageRAG(BaseRAGExample):
+    """RAG example for iMessage conversation history."""
+
+    def __init__(self):
+        super().__init__(
+            name="iMessage",
+            description="RAG on your iMessage conversation history",
+            default_index_name="imessage_index",
+        )
+
+    def _add_specific_arguments(self, parser):
+        """Add iMessage-specific arguments."""
+        imessage_group = parser.add_argument_group("iMessage Parameters")
+        imessage_group.add_argument(
+            "--db-path",
+            type=str,
+            default=None,
+            help="Path to iMessage chat.db file (default: ~/Library/Messages/chat.db)",
+        )
+        imessage_group.add_argument(
+            "--concatenate-conversations",
+            action="store_true",
+            default=True,
+            help="Concatenate messages within conversations for better context (default: True)",
+        )
+        imessage_group.add_argument(
+            "--no-concatenate-conversations",
+            action="store_true",
+            help="Process each message individually instead of concatenating by conversation",
+        )
+        imessage_group.add_argument(
+            "--chunk-size",
+            type=int,
+            default=1000,
+            help="Maximum characters per text chunk (default: 1000)",
+        )
+        imessage_group.add_argument(
+            "--chunk-overlap",
+            type=int,
+            default=200,
+            help="Overlap between text chunks (default: 200)",
+        )
+
+    async def load_data(self, args) -> list[str]:
+        """Load iMessage history and convert to text chunks."""
+        print("Loading iMessage conversation history...")
+
+        # Determine concatenation setting
+        concatenate = args.concatenate_conversations and not args.no_concatenate_conversations
+
+        # Initialize iMessage reader
+        reader = IMessageReader(concatenate_conversations=concatenate)
+
+        # Load documents
+        try:
+            if args.db_path:
+                # Use custom database path
+                db_dir = str(Path(args.db_path).parent)
+                documents = reader.load_data(input_dir=db_dir)
+            else:
+                # Use default macOS location
+                documents = reader.load_data()
+
+        except Exception as e:
+            print(f"Error loading iMessage data: {e}")
+            print("\nTroubleshooting tips:")
+            print("1. Make sure you have granted Full Disk Access to your terminal/IDE")
+            print("2. Check that the iMessage database exists at ~/Library/Messages/chat.db")
+            print("3. Try specifying a custom path with --db-path if you have a backup")
+            return []
+
+        if not documents:
+            print("No iMessage conversations found!")
+            return []
+
+        print(f"Loaded {len(documents)} iMessage documents")
+
+        # Show some statistics
+        total_messages = sum(doc.metadata.get("message_count", 1) for doc in documents)
+        print(f"Total messages: {total_messages}")
+
+        if concatenate:
+            # Show chat statistics
+            chat_names = [doc.metadata.get("chat_name", "Unknown") for doc in documents]
+            unique_chats = len(set(chat_names))
+            print(f"Unique conversations: {unique_chats}")
+
+        # Convert to text chunks
+        all_texts = create_text_chunks(
+            documents,
+            chunk_size=args.chunk_size,
+            chunk_overlap=args.chunk_overlap,
+        )
+
+        # Apply max_items limit if specified
+        if args.max_items > 0:
+            all_texts = all_texts[: args.max_items]
+            print(f"Limited to {len(all_texts)} text chunks (max_items={args.max_items})")
+
+        return all_texts
+
+
+async def main():
+    """Main entry point."""
+    app = IMessageRAG()
+    await app.run()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

apps/multimodal/vision-based-pdf-multi-vector/README.md

@@ -0,0 +1,113 @@
+## Vision-based PDF Multi-Vector Demos (macOS/MPS)
+
+This folder contains two demos to index PDF pages as images and run multi-vector retrieval with ColPali/ColQwen2, plus optional similarity map visualization and answer generation.
+
+### What you’ll run
+- `multi-vector-leann-paper-example.py`: local PDF → pages → embed → build HNSW index → search.
+- `multi-vector-leann-similarity-map.py`: HF dataset (default) or local pages → embed → index → retrieve → similarity maps → optional Qwen-VL answer.
+
+## Prerequisites (macOS)
+
+### 1) Homebrew poppler (for pdf2image)
+```bash
+brew install poppler
+which pdfinfo && pdfinfo -v
+```
+
+### 2) Python environment
+Use uv (recommended) or pip. Python 3.9+.
+
+Using uv:
+```bash
+uv pip install \
+  colpali_engine \
+  pdf2image \
+  pillow \
+  matplotlib qwen_vl_utils \
+  einops \
+  seaborn
+```
+
+Notes:
+- On first run, models download from Hugging Face. Login/config if needed.
+- The scripts auto-select device: CUDA > MPS > CPU. Verify MPS:
+```bash
+python -c "import torch; print('MPS available:', bool(getattr(torch.backends, 'mps', None) and torch.backends.mps.is_available()))"
+```
+
+## Run the demos
+
+### A) Local PDF example
+Converts a local PDF into page images, embeds them, builds an index, and searches.
+
+```bash
+cd apps/multimodal/vision-based-pdf-multi-vector
+# If you don't have the sample PDF locally, download it (ignored by Git)
+mkdir -p pdfs
+curl -L -o pdfs/2004.12832v2.pdf https://arxiv.org/pdf/2004.12832.pdf
+ls pdfs/2004.12832v2.pdf
+# Ensure output dir exists
+mkdir -p pages
+python multi-vector-leann-paper-example.py
+```
+Expected:
+- Page images in `pages/`.
+- Console prints like `Using device=mps, dtype=...` and retrieved file paths for queries.
+
+To use your own PDF: edit `pdf_path` near the top of the script.
+
+### B) Similarity map + answer demo
+Uses HF dataset `weaviate/arXiv-AI-papers-multi-vector` by default; can switch to local pages.
+
+```bash
+cd apps/multimodal/vision-based-pdf-multi-vector
+python multi-vector-leann-similarity-map.py
+```
+Artifacts (when enabled):
+- Retrieved pages: `./figures/retrieved_page_rank{K}.png`
+- Similarity maps: `./figures/similarity_map_rank{K}.png`
+
+Key knobs in the script (top of file):
+- `QUERY`: your question
+- `MODEL`: `"colqwen2"` or `"colpali"`
+- `USE_HF_DATASET`: set `False` to use local pages
+- `PDF`, `PAGES_DIR`: for local mode
+- `INDEX_PATH`, `TOPK`, `FIRST_STAGE_K`, `REBUILD_INDEX`
+- `SIMILARITY_MAP`, `SIM_TOKEN_IDX`, `SIM_OUTPUT`
+- `ANSWER`, `MAX_NEW_TOKENS` (Qwen-VL)
+
+## Troubleshooting
+- pdf2image errors on macOS: ensure `brew install poppler` and `pdfinfo` works in terminal.
+- Slow or OOM on MPS: reduce dataset size (e.g., set `MAX_DOCS`) or switch to CPU.
+- NaNs on MPS: keep fp32 on MPS (default in similarity-map script); avoid fp16 there.
+- First-run model downloads can be large; ensure network access (HF mirrors if needed).
+
+## Notes
+- Index files are under `./indexes/`. Delete or set `REBUILD_INDEX=True` to rebuild.
+- For local PDFs, page images go to `./pages/`.
+
+
+### Retrieval and Visualization Example
+
+Example settings in `multi-vector-leann-similarity-map.py`:
+- `QUERY = "How does DeepSeek-V2 compare against the LLaMA family of LLMs?"`
+- `SIMILARITY_MAP = True` (to generate heatmaps)
+- `TOPK = 1` (save the top retrieved page and its similarity map)
+
+Run:
+```bash
+cd apps/multimodal/vision-based-pdf-multi-vector
+python multi-vector-leann-similarity-map.py
+```
+
+Outputs (by default):
+- Retrieved page: `./figures/retrieved_page_rank1.png`
+- Similarity map: `./figures/similarity_map_rank1.png`
+
+Sample visualization (example result, and the query is "QUERY = "How does Vim model performance and efficiency compared to other models?"
+"):
+![Similarity map example](fig/image.png)
+
+Notes:
+- Set `SIM_TOKEN_IDX` to visualize a specific token index; set `-1` to auto-select the most salient token.
+- If you change `SIM_OUTPUT` to a file path (e.g., `./figures/my_map.png`), multiple ranks are saved as `my_map_rank{K}.png`.

apps/multimodal/vision-based-pdf-multi-vector/leann_multi_vector.py

@@ -0,0 +1,182 @@
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+
+
+def _ensure_repo_paths_importable(current_file: str) -> None:
+    _repo_root = Path(current_file).resolve().parents[3]
+    _leann_core_src = _repo_root / "packages" / "leann-core" / "src"
+    _leann_hnsw_pkg = _repo_root / "packages" / "leann-backend-hnsw"
+    if str(_leann_core_src) not in sys.path:
+        sys.path.append(str(_leann_core_src))
+    if str(_leann_hnsw_pkg) not in sys.path:
+        sys.path.append(str(_leann_hnsw_pkg))
+
+
+_ensure_repo_paths_importable(__file__)
+
+from leann_backend_hnsw.hnsw_backend import HNSWBuilder, HNSWSearcher  # noqa: E402
+
+
+class LeannMultiVector:
+    def __init__(
+        self,
+        index_path: str,
+        dim: int = 128,
+        distance_metric: str = "mips",
+        m: int = 16,
+        ef_construction: int = 500,
+        is_compact: bool = False,
+        is_recompute: bool = False,
+        embedding_model_name: str = "colvision",
+    ) -> None:
+        self.index_path = index_path
+        self.dim = dim
+        self.embedding_model_name = embedding_model_name
+        self._pending_items: list[dict] = []
+        self._backend_kwargs = {
+            "distance_metric": distance_metric,
+            "M": m,
+            "efConstruction": ef_construction,
+            "is_compact": is_compact,
+            "is_recompute": is_recompute,
+        }
+        self._labels_meta: list[dict] = []
+
+    def _meta_dict(self) -> dict:
+        return {
+            "version": "1.0",
+            "backend_name": "hnsw",
+            "embedding_model": self.embedding_model_name,
+            "embedding_mode": "custom",
+            "dimensions": self.dim,
+            "backend_kwargs": self._backend_kwargs,
+            "is_compact": self._backend_kwargs.get("is_compact", True),
+            "is_pruned": self._backend_kwargs.get("is_compact", True)
+            and self._backend_kwargs.get("is_recompute", True),
+        }
+
+    def create_collection(self) -> None:
+        path = Path(self.index_path)
+        path.parent.mkdir(parents=True, exist_ok=True)
+
+    def insert(self, data: dict) -> None:
+        self._pending_items.append(
+            {
+                "doc_id": int(data["doc_id"]),
+                "filepath": data.get("filepath", ""),
+                "colbert_vecs": [np.asarray(v, dtype=np.float32) for v in data["colbert_vecs"]],
+            }
+        )
+
+    def _labels_path(self) -> Path:
+        index_path_obj = Path(self.index_path)
+        return index_path_obj.parent / f"{index_path_obj.name}.labels.json"
+
+    def _meta_path(self) -> Path:
+        index_path_obj = Path(self.index_path)
+        return index_path_obj.parent / f"{index_path_obj.name}.meta.json"
+
+    def create_index(self) -> None:
+        if not self._pending_items:
+            return
+
+        embeddings: list[np.ndarray] = []
+        labels_meta: list[dict] = []
+
+        for item in self._pending_items:
+            doc_id = int(item["doc_id"])
+            filepath = item.get("filepath", "")
+            colbert_vecs = item["colbert_vecs"]
+            for seq_id, vec in enumerate(colbert_vecs):
+                vec_np = np.asarray(vec, dtype=np.float32)
+                embeddings.append(vec_np)
+                labels_meta.append(
+                    {
+                        "id": f"{doc_id}:{seq_id}",
+                        "doc_id": doc_id,
+                        "seq_id": int(seq_id),
+                        "filepath": filepath,
+                    }
+                )
+
+        if not embeddings:
+            return
+
+        embeddings_np = np.vstack(embeddings).astype(np.float32)
+        # print shape of embeddings_np
+        print(embeddings_np.shape)
+
+        builder = HNSWBuilder(**{**self._backend_kwargs, "dimensions": self.dim})
+        ids = [str(i) for i in range(embeddings_np.shape[0])]
+        builder.build(embeddings_np, ids, self.index_path)
+
+        import json as _json
+
+        with open(self._meta_path(), "w", encoding="utf-8") as f:
+            _json.dump(self._meta_dict(), f, indent=2)
+        with open(self._labels_path(), "w", encoding="utf-8") as f:
+            _json.dump(labels_meta, f)
+
+        self._labels_meta = labels_meta
+
+    def _load_labels_meta_if_needed(self) -> None:
+        if self._labels_meta:
+            return
+        labels_path = self._labels_path()
+        if labels_path.exists():
+            import json as _json
+
+            with open(labels_path, encoding="utf-8") as f:
+                self._labels_meta = _json.load(f)
+
+    def search(
+        self, data: np.ndarray, topk: int, first_stage_k: int = 50
+    ) -> list[tuple[float, int]]:
+        if data.ndim == 1:
+            data = data.reshape(1, -1)
+        if data.dtype != np.float32:
+            data = data.astype(np.float32)
+
+        self._load_labels_meta_if_needed()
+
+        searcher = HNSWSearcher(self.index_path, meta=self._meta_dict())
+        raw = searcher.search(
+            data,
+            first_stage_k,
+            recompute_embeddings=False,
+            complexity=128,
+            beam_width=1,
+            prune_ratio=0.0,
+            batch_size=0,
+        )
+
+        labels = raw.get("labels")
+        distances = raw.get("distances")
+        if labels is None or distances is None:
+            return []
+
+        doc_scores: dict[int, float] = {}
+        B = len(labels)
+        for b in range(B):
+            per_doc_best: dict[int, float] = {}
+            for k, sid in enumerate(labels[b]):
+                try:
+                    idx = int(sid)
+                except Exception:
+                    continue
+                if 0 <= idx < len(self._labels_meta):
+                    doc_id = int(self._labels_meta[idx]["doc_id"])  # type: ignore[index]
+                else:
+                    continue
+                score = float(distances[b][k])
+                if (doc_id not in per_doc_best) or (score > per_doc_best[doc_id]):
+                    per_doc_best[doc_id] = score
+            for doc_id, best_score in per_doc_best.items():
+                doc_scores[doc_id] = doc_scores.get(doc_id, 0.0) + best_score
+
+        scores = sorted(((v, k) for k, v in doc_scores.items()), key=lambda x: x[0], reverse=True)
+        return scores[:topk] if len(scores) >= topk else scores

apps/multimodal/vision-based-pdf-multi-vector/multi-vector-leann-paper-example.py

@@ -0,0 +1,112 @@
+# pip install pdf2image
+# pip install pymilvus
+# pip install colpali_engine
+# pip install tqdm
+# pip install pillow
+
+import os
+import re
+import sys
+from pathlib import Path
+from typing import cast
+
+from PIL import Image
+from tqdm import tqdm
+
+# Ensure local leann packages are importable before importing them
+_repo_root = Path(__file__).resolve().parents[3]
+_leann_core_src = _repo_root / "packages" / "leann-core" / "src"
+_leann_hnsw_pkg = _repo_root / "packages" / "leann-backend-hnsw"
+if str(_leann_core_src) not in sys.path:
+    sys.path.append(str(_leann_core_src))
+if str(_leann_hnsw_pkg) not in sys.path:
+    sys.path.append(str(_leann_hnsw_pkg))
+
+
+import torch
+from colpali_engine.models import ColPali
+from colpali_engine.models.paligemma.colpali.processing_colpali import ColPaliProcessor
+from colpali_engine.utils.torch_utils import ListDataset, get_torch_device
+from torch.utils.data import DataLoader
+
+# Auto-select device: CUDA > MPS (mac) > CPU
+_device_str = (
+    "cuda"
+    if torch.cuda.is_available()
+    else (
+        "mps"
+        if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available()
+        else "cpu"
+    )
+)
+device = get_torch_device(_device_str)
+# Prefer fp16 on GPU/MPS, bfloat16 on CPU
+_dtype = torch.float16 if _device_str in ("cuda", "mps") else torch.bfloat16
+model_name = "vidore/colpali-v1.2"
+
+model = ColPali.from_pretrained(
+    model_name,
+    torch_dtype=_dtype,
+    device_map=device,
+).eval()
+print(f"Using device={_device_str}, dtype={_dtype}")
+
+queries = [
+    "How to end-to-end retrieval with ColBert",
+    "Where is ColBERT performance Table, including text representation results?",
+]
+
+processor = cast(ColPaliProcessor, ColPaliProcessor.from_pretrained(model_name))
+
+dataloader = DataLoader(
+    dataset=ListDataset[str](queries),
+    batch_size=1,
+    shuffle=False,
+    collate_fn=lambda x: processor.process_queries(x),
+)
+
+qs: list[torch.Tensor] = []
+for batch_query in dataloader:
+    with torch.no_grad():
+        batch_query = {k: v.to(model.device) for k, v in batch_query.items()}
+        embeddings_query = model(**batch_query)
+    qs.extend(list(torch.unbind(embeddings_query.to("cpu"))))
+print(qs[0].shape)
+# %%
+page_filenames = sorted(os.listdir("./pages"), key=lambda n: int(re.search(r"\d+", n).group()))
+images = [Image.open(os.path.join("./pages", name)) for name in page_filenames]
+
+dataloader = DataLoader(
+    dataset=ListDataset[str](images),
+    batch_size=1,
+    shuffle=False,
+    collate_fn=lambda x: processor.process_images(x),
+)
+
+ds: list[torch.Tensor] = []
+for batch_doc in tqdm(dataloader):
+    with torch.no_grad():
+        batch_doc = {k: v.to(model.device) for k, v in batch_doc.items()}
+        embeddings_doc = model(**batch_doc)
+    ds.extend(list(torch.unbind(embeddings_doc.to("cpu"))))
+
+print(ds[0].shape)
+
+# %%
+# Build HNSW index via LeannRetriever primitives and run search
+index_path = "./indexes/colpali.leann"
+retriever = LeannRetriever(index_path=index_path, dim=int(ds[0].shape[-1]))
+retriever.create_collection()
+filepaths = [os.path.join("./pages", name) for name in page_filenames]
+for i in range(len(filepaths)):
+    data = {
+        "colbert_vecs": ds[i].float().numpy(),
+        "doc_id": i,
+        "filepath": filepaths[i],
+    }
+    retriever.insert(data)
+retriever.create_index()
+for query in qs:
+    query_np = query.float().numpy()
+    result = retriever.search(query_np, topk=1)
+    print(filepaths[result[0][1]])

apps/multimodal/vision-based-pdf-multi-vector/multi-vector-leann-similarity-map.py

@@ -0,0 +1,477 @@
+## Jupyter-style notebook script
+# %%
+# uv pip install matplotlib qwen_vl_utils
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Any, Optional, cast
+
+from PIL import Image
+from tqdm import tqdm
+
+
+def _ensure_repo_paths_importable(current_file: str) -> None:
+    """Make local leann packages importable without installing (mirrors multi-vector-leann.py)."""
+    _repo_root = Path(current_file).resolve().parents[3]
+    _leann_core_src = _repo_root / "packages" / "leann-core" / "src"
+    _leann_hnsw_pkg = _repo_root / "packages" / "leann-backend-hnsw"
+    if str(_leann_core_src) not in sys.path:
+        sys.path.append(str(_leann_core_src))
+    if str(_leann_hnsw_pkg) not in sys.path:
+        sys.path.append(str(_leann_hnsw_pkg))
+
+
+_ensure_repo_paths_importable(__file__)
+
+from leann_multi_vector import LeannMultiVector  # noqa: E402
+
+# %%
+# Config
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+QUERY = "How does DeepSeek-V2 compare against the LLaMA family of LLMs?"
+MODEL: str = "colqwen2"  # "colpali" or "colqwen2"
+
+# Data source: set to True to use the Hugging Face dataset example (recommended)
+USE_HF_DATASET: bool = True
+DATASET_NAME: str = "weaviate/arXiv-AI-papers-multi-vector"
+DATASET_SPLIT: str = "train"
+MAX_DOCS: Optional[int] = None  # limit number of pages to index; None = all
+
+# Local pages (used when USE_HF_DATASET == False)
+PDF: Optional[str] = None  # e.g., "./pdfs/2004.12832v2.pdf"
+PAGES_DIR: str = "./pages"
+
+# Index + retrieval settings
+INDEX_PATH: str = "./indexes/colvision.leann"
+TOPK: int = 1
+FIRST_STAGE_K: int = 500
+REBUILD_INDEX: bool = False
+
+# Artifacts
+SAVE_TOP_IMAGE: Optional[str] = "./figures/retrieved_page.png"
+SIMILARITY_MAP: bool = True
+SIM_TOKEN_IDX: int = 13  # -1 means auto-select the most salient token
+SIM_OUTPUT: str = "./figures/similarity_map.png"
+ANSWER: bool = True
+MAX_NEW_TOKENS: int = 128
+
+
+# %%
+# Helpers
+def _natural_sort_key(name: str) -> int:
+    m = re.search(r"\d+", name)
+    return int(m.group()) if m else 0
+
+
+def _load_images_from_dir(pages_dir: str) -> tuple[list[str], list[Image.Image]]:
+    filenames = [n for n in os.listdir(pages_dir) if n.lower().endswith((".png", ".jpg", ".jpeg"))]
+    filenames = sorted(filenames, key=_natural_sort_key)
+    filepaths = [os.path.join(pages_dir, n) for n in filenames]
+    images = [Image.open(p) for p in filepaths]
+    return filepaths, images
+
+
+def _maybe_convert_pdf_to_images(pdf_path: Optional[str], pages_dir: str, dpi: int = 200) -> None:
+    if not pdf_path:
+        return
+    os.makedirs(pages_dir, exist_ok=True)
+    try:
+        from pdf2image import convert_from_path
+    except Exception as e:
+        raise RuntimeError(
+            "pdf2image is required to convert PDF to images. Install via pip install pdf2image"
+        ) from e
+    images = convert_from_path(pdf_path, dpi=dpi)
+    for i, image in enumerate(images):
+        image.save(os.path.join(pages_dir, f"page_{i + 1}.png"), "PNG")
+
+
+def _select_device_and_dtype():
+    import torch
+    from colpali_engine.utils.torch_utils import get_torch_device
+
+    device_str = (
+        "cuda"
+        if torch.cuda.is_available()
+        else (
+            "mps"
+            if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available()
+            else "cpu"
+        )
+    )
+    device = get_torch_device(device_str)
+    # Stable dtype selection to avoid NaNs:
+    # - CUDA: prefer bfloat16 if supported, else float16
+    # - MPS: use float32 (fp16 on MPS can produce NaNs in some ops)
+    # - CPU: float32
+    if device_str == "cuda":
+        dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+        try:
+            torch.backends.cuda.matmul.allow_tf32 = True  # Better stability/perf on Ampere+
+        except Exception:
+            pass
+    elif device_str == "mps":
+        dtype = torch.float32
+    else:
+        dtype = torch.float32
+    return device_str, device, dtype
+
+
+def _load_colvision(model_choice: str):
+    import torch
+    from colpali_engine.models import ColPali, ColQwen2, ColQwen2Processor
+    from colpali_engine.models.paligemma.colpali.processing_colpali import ColPaliProcessor
+    from transformers.utils.import_utils import is_flash_attn_2_available
+
+    device_str, device, dtype = _select_device_and_dtype()
+
+    if model_choice == "colqwen2":
+        model_name = "vidore/colqwen2-v1.0"
+        # On CPU/MPS we must avoid flash-attn and stay eager; on CUDA prefer flash-attn if available
+        attn_implementation = (
+            "flash_attention_2"
+            if (device_str == "cuda" and is_flash_attn_2_available())
+            else "eager"
+        )
+        model = ColQwen2.from_pretrained(
+            model_name,
+            torch_dtype=torch.bfloat16,
+            device_map=device,
+            attn_implementation=attn_implementation,
+        ).eval()
+        processor = ColQwen2Processor.from_pretrained(model_name)
+    else:
+        model_name = "vidore/colpali-v1.2"
+        model = ColPali.from_pretrained(
+            model_name,
+            torch_dtype=torch.bfloat16,
+            device_map=device,
+        ).eval()
+        processor = cast(ColPaliProcessor, ColPaliProcessor.from_pretrained(model_name))
+
+    return model_name, model, processor, device_str, device, dtype
+
+
+def _embed_images(model, processor, images: list[Image.Image]) -> list[Any]:
+    import torch
+    from colpali_engine.utils.torch_utils import ListDataset
+    from torch.utils.data import DataLoader
+
+    # Ensure deterministic eval and autocast for stability
+    model.eval()
+
+    dataloader = DataLoader(
+        dataset=ListDataset[Image.Image](images),
+        batch_size=1,
+        shuffle=False,
+        collate_fn=lambda x: processor.process_images(x),
+    )
+
+    doc_vecs: list[Any] = []
+    for batch_doc in tqdm(dataloader, desc="Embedding images"):
+        with torch.no_grad():
+            batch_doc = {k: v.to(model.device) for k, v in batch_doc.items()}
+            # autocast on CUDA for bf16/fp16; on CPU/MPS stay in fp32
+            if model.device.type == "cuda":
+                with torch.autocast(
+                    device_type="cuda",
+                    dtype=model.dtype if model.dtype.is_floating_point else torch.bfloat16,
+                ):
+                    embeddings_doc = model(**batch_doc)
+            else:
+                embeddings_doc = model(**batch_doc)
+        doc_vecs.extend(list(torch.unbind(embeddings_doc.to("cpu"))))
+    return doc_vecs
+
+
+def _embed_queries(model, processor, queries: list[str]) -> list[Any]:
+    import torch
+    from colpali_engine.utils.torch_utils import ListDataset
+    from torch.utils.data import DataLoader
+
+    model.eval()
+
+    dataloader = DataLoader(
+        dataset=ListDataset[str](queries),
+        batch_size=1,
+        shuffle=False,
+        collate_fn=lambda x: processor.process_queries(x),
+    )
+
+    q_vecs: list[Any] = []
+    for batch_query in tqdm(dataloader, desc="Embedding queries"):
+        with torch.no_grad():
+            batch_query = {k: v.to(model.device) for k, v in batch_query.items()}
+            if model.device.type == "cuda":
+                with torch.autocast(
+                    device_type="cuda",
+                    dtype=model.dtype if model.dtype.is_floating_point else torch.bfloat16,
+                ):
+                    embeddings_query = model(**batch_query)
+            else:
+                embeddings_query = model(**batch_query)
+        q_vecs.extend(list(torch.unbind(embeddings_query.to("cpu"))))
+    return q_vecs
+
+
+def _build_index(index_path: str, doc_vecs: list[Any], filepaths: list[str]) -> LeannMultiVector:
+    dim = int(doc_vecs[0].shape[-1])
+    retriever = LeannMultiVector(index_path=index_path, dim=dim)
+    retriever.create_collection()
+    for i, vec in enumerate(doc_vecs):
+        data = {
+            "colbert_vecs": vec.float().numpy(),
+            "doc_id": i,
+            "filepath": filepaths[i],
+        }
+        retriever.insert(data)
+    retriever.create_index()
+    return retriever
+
+
+def _load_retriever_if_index_exists(index_path: str, dim: int) -> Optional[LeannMultiVector]:
+    index_base = Path(index_path)
+    # Rough heuristic: index dir exists AND meta+labels files exist
+    meta = index_base.parent / f"{index_base.name}.meta.json"
+    labels = index_base.parent / f"{index_base.name}.labels.json"
+    if index_base.exists() and meta.exists() and labels.exists():
+        return LeannMultiVector(index_path=index_path, dim=dim)
+    return None
+
+
+def _generate_similarity_map(
+    model,
+    processor,
+    image: Image.Image,
+    query: str,
+    token_idx: Optional[int] = None,
+    output_path: Optional[str] = None,
+) -> tuple[int, float]:
+    import torch
+    from colpali_engine.interpretability import (
+        get_similarity_maps_from_embeddings,
+        plot_similarity_map,
+    )
+
+    batch_images = processor.process_images([image]).to(model.device)
+    batch_queries = processor.process_queries([query]).to(model.device)
+
+    with torch.no_grad():
+        image_embeddings = model.forward(**batch_images)
+        query_embeddings = model.forward(**batch_queries)
+
+    n_patches = processor.get_n_patches(
+        image_size=image.size,
+        spatial_merge_size=getattr(model, "spatial_merge_size", None),
+    )
+    image_mask = processor.get_image_mask(batch_images)
+
+    batched_similarity_maps = get_similarity_maps_from_embeddings(
+        image_embeddings=image_embeddings,
+        query_embeddings=query_embeddings,
+        n_patches=n_patches,
+        image_mask=image_mask,
+    )
+
+    similarity_maps = batched_similarity_maps[0]
+
+    # Determine token index if not provided: choose the token with highest max score
+    if token_idx is None:
+        per_token_max = similarity_maps.view(similarity_maps.shape[0], -1).max(dim=1).values
+        token_idx = int(per_token_max.argmax().item())
+
+    max_sim_score = similarity_maps[token_idx, :, :].max().item()
+
+    if output_path:
+        import matplotlib.pyplot as plt
+
+        fig, ax = plot_similarity_map(
+            image=image,
+            similarity_map=similarity_maps[token_idx],
+            figsize=(14, 14),
+            show_colorbar=False,
+        )
+        ax.set_title(f"Token #{token_idx}. MaxSim score: {max_sim_score:.2f}", fontsize=12)
+        os.makedirs(os.path.dirname(output_path), exist_ok=True)
+        plt.savefig(output_path, bbox_inches="tight")
+        plt.close(fig)
+
+    return token_idx, float(max_sim_score)
+
+
+class QwenVL:
+    def __init__(self, device: str):
+        from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+        from transformers.utils.import_utils import is_flash_attn_2_available
+
+        attn_implementation = "flash_attention_2" if is_flash_attn_2_available() else "eager"
+        self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2.5-VL-3B-Instruct",
+            torch_dtype="auto",
+            device_map=device,
+            attn_implementation=attn_implementation,
+        )
+
+        min_pixels = 256 * 28 * 28
+        max_pixels = 1280 * 28 * 28
+        self.processor = AutoProcessor.from_pretrained(
+            "Qwen/Qwen2.5-VL-3B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels
+        )
+
+    def answer(self, query: str, images: list[Image.Image], max_new_tokens: int = 128) -> str:
+        import base64
+        from io import BytesIO
+
+        from qwen_vl_utils import process_vision_info
+
+        content = []
+        for img in images:
+            buffer = BytesIO()
+            img.save(buffer, format="jpeg")
+            img_base64 = base64.b64encode(buffer.getvalue()).decode("utf-8")
+            content.append({"type": "image", "image": f"data:image;base64,{img_base64}"})
+        content.append({"type": "text", "text": query})
+        messages = [{"role": "user", "content": content}]
+
+        text = self.processor.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=True
+        )
+        image_inputs, video_inputs = process_vision_info(messages)
+        inputs = self.processor(
+            text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt"
+        )
+        inputs = inputs.to(self.model.device)
+
+        generated_ids = self.model.generate(**inputs, max_new_tokens=max_new_tokens)
+        generated_ids_trimmed = [
+            out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+        ]
+        return self.processor.batch_decode(
+            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )[0]
+
+
+# %%
+
+# Step 1: Prepare data
+if USE_HF_DATASET:
+    from datasets import load_dataset
+
+    dataset = load_dataset(DATASET_NAME, split=DATASET_SPLIT)
+    N = len(dataset) if MAX_DOCS is None else min(MAX_DOCS, len(dataset))
+    filepaths: list[str] = []
+    images: list[Image.Image] = []
+    for i in tqdm(range(N), desc="Loading dataset", total=N ):
+        p = dataset[i]
+        # Compose a descriptive identifier for printing later
+        identifier = f"arXiv:{p['paper_arxiv_id']}|title:{p['paper_title']}|page:{int(p['page_number'])}|id:{p['page_id']}"
+        print(identifier)
+        filepaths.append(identifier)
+        images.append(p["page_image"])  # PIL Image
+else:
+    _maybe_convert_pdf_to_images(PDF, PAGES_DIR)
+    filepaths, images = _load_images_from_dir(PAGES_DIR)
+    if not images:
+        raise RuntimeError(
+            f"No images found in {PAGES_DIR}. Provide PDF path in PDF variable or ensure images exist."
+        )
+
+
+# %%
+# Step 2: Load model and processor
+model_name, model, processor, device_str, device, dtype = _load_colvision(MODEL)
+print(f"Using model={model_name}, device={device_str}, dtype={dtype}")
+
+
+# %%
+
+# %%
+# Step 3: Build or load index
+retriever: Optional[LeannMultiVector] = None
+if not REBUILD_INDEX:
+    try:
+        one_vec = _embed_images(model, processor, [images[0]])[0]
+        retriever = _load_retriever_if_index_exists(INDEX_PATH, dim=int(one_vec.shape[-1]))
+    except Exception:
+        retriever = None
+
+if retriever is None:
+    doc_vecs = _embed_images(model, processor, images)
+    retriever = _build_index(INDEX_PATH, doc_vecs, filepaths)
+
+
+# %%
+# Step 4: Embed query and search
+q_vec = _embed_queries(model, processor, [QUERY])[0]
+results = retriever.search(q_vec.float().numpy(), topk=TOPK, first_stage_k=FIRST_STAGE_K)
+if not results:
+    print("No results found.")
+else:
+    print(f'Top {len(results)} results for query: "{QUERY}"')
+    top_images: list[Image.Image] = []
+    for rank, (score, doc_id) in enumerate(results, start=1):
+        path = filepaths[doc_id]
+        # For HF dataset, path is a descriptive identifier, not a real file path
+        print(f"{rank}) MaxSim: {score:.4f}, Page: {path}")
+        top_images.append(images[doc_id])
+
+    if SAVE_TOP_IMAGE:
+        from pathlib import Path as _Path
+
+        base = _Path(SAVE_TOP_IMAGE)
+        base.parent.mkdir(parents=True, exist_ok=True)
+        for rank, img in enumerate(top_images[:TOPK], start=1):
+            if base.suffix:
+                out_path = base.parent / f"{base.stem}_rank{rank}{base.suffix}"
+            else:
+                out_path = base / f"retrieved_page_rank{rank}.png"
+            img.save(str(out_path))
+            print(f"Saved retrieved page (rank {rank}) to: {out_path}")
+
+## TODO stange results of second page of DeepSeek-V2 rather than the first page
+
+# %%
+# Step 5: Similarity maps for top-K results
+if results and SIMILARITY_MAP:
+    token_idx = None if SIM_TOKEN_IDX < 0 else int(SIM_TOKEN_IDX)
+    from pathlib import Path as _Path
+
+    output_base = _Path(SIM_OUTPUT) if SIM_OUTPUT else None
+    for rank, img in enumerate(top_images[:TOPK], start=1):
+        if output_base:
+            if output_base.suffix:
+                out_dir = output_base.parent
+                out_name = f"{output_base.stem}_rank{rank}{output_base.suffix}"
+                out_path = str(out_dir / out_name)
+            else:
+                out_dir = output_base
+                out_dir.mkdir(parents=True, exist_ok=True)
+                out_path = str(out_dir / f"similarity_map_rank{rank}.png")
+        else:
+            out_path = None
+        chosen_idx, max_sim = _generate_similarity_map(
+            model=model,
+            processor=processor,
+            image=img,
+            query=QUERY,
+            token_idx=token_idx,
+            output_path=out_path,
+        )
+        if out_path:
+            print(
+                f"Saved similarity map for rank {rank}, token #{chosen_idx} (max={max_sim:.2f}) to: {out_path}"
+            )
+        else:
+            print(
+                f"Computed similarity map for rank {rank}, token #{chosen_idx} (max={max_sim:.2f})"
+            )
+
+
+# %%
+# Step 6: Optional answer generation
+if results and ANSWER:
+    qwen = QwenVL(device=device_str)
+    response = qwen.answer(QUERY, top_images[:TOPK], max_new_tokens=MAX_NEW_TOKENS)
+    print("\nAnswer:")
+    print(response)

apps/semantic_file_search/leann-plus-temporal-search.py

@@ -0,0 +1,183 @@
+#!/usr/bin/env python3
+import re
+import sys
+from datetime import datetime, timedelta
+from pathlib import Path
+
+from leann import LeannSearcher
+
+INDEX_PATH = str(Path("./").resolve() / "demo.leann")
+
+
+class TimeParser:
+    def __init__(self):
+        # Main pattern: captures optional fuzzy modifier, number, unit, and optional "ago"
+        self.pattern = r"(?:(around|about|roughly|approximately)\s+)?(\d+)\s+(hour|day|week|month|year)s?(?:\s+ago)?"
+
+        # Compile for performance
+        self.regex = re.compile(self.pattern, re.IGNORECASE)
+
+        # Stop words to remove before regex parsing
+        self.stop_words = {
+            "in",
+            "at",
+            "of",
+            "by",
+            "as",
+            "me",
+            "the",
+            "a",
+            "an",
+            "and",
+            "any",
+            "find",
+            "search",
+            "list",
+            "ago",
+            "back",
+            "past",
+            "earlier",
+        }
+
+    def clean_text(self, text):
+        """Remove stop words from text"""
+        words = text.split()
+        cleaned = " ".join(word for word in words if word.lower() not in self.stop_words)
+        return cleaned
+
+    def parse(self, text):
+        """Extract all time expressions from text"""
+        # Clean text first
+        cleaned_text = self.clean_text(text)
+
+        matches = []
+        for match in self.regex.finditer(cleaned_text):
+            fuzzy = match.group(1)  # "around", "about", etc.
+            number = int(match.group(2))
+            unit = match.group(3).lower()
+
+            matches.append(
+                {
+                    "full_match": match.group(0),
+                    "fuzzy": bool(fuzzy),
+                    "number": number,
+                    "unit": unit,
+                    "range": self.calculate_range(number, unit, bool(fuzzy)),
+                }
+            )
+
+        return matches
+
+    def calculate_range(self, number, unit, is_fuzzy):
+        """Convert to actual datetime range and return ISO format strings"""
+        units = {
+            "hour": timedelta(hours=number),
+            "day": timedelta(days=number),
+            "week": timedelta(weeks=number),
+            "month": timedelta(days=number * 30),
+            "year": timedelta(days=number * 365),
+        }
+
+        delta = units[unit]
+        now = datetime.now()
+        target = now - delta
+
+        if is_fuzzy:
+            buffer = delta * 0.2  # 20% buffer for fuzzy
+            start = (target - buffer).isoformat()
+            end = (target + buffer).isoformat()
+        else:
+            start = target.isoformat()
+            end = now.isoformat()
+
+        return (start, end)
+
+
+def search_files(query, top_k=15):
+    """Search the index and return results"""
+    # Parse time expressions
+    parser = TimeParser()
+    time_matches = parser.parse(query)
+
+    # Remove time expressions from query for semantic search
+    clean_query = query
+    if time_matches:
+        for match in time_matches:
+            clean_query = clean_query.replace(match["full_match"], "").strip()
+
+    # Check if clean_query is less than 4 characters
+    if len(clean_query) < 4:
+        print("Error: add more input for accurate results.")
+        return
+
+    # Single query to vector DB
+    searcher = LeannSearcher(INDEX_PATH)
+    results = searcher.search(
+        clean_query if clean_query else query, top_k=top_k, recompute_embeddings=False
+    )
+
+    # Filter by time if time expression found
+    if time_matches:
+        time_range = time_matches[0]["range"]  # Use first time expression
+        start_time, end_time = time_range
+
+        filtered_results = []
+        for result in results:
+            # Access metadata attribute directly (not .get())
+            metadata = result.metadata if hasattr(result, "metadata") else {}
+
+            if metadata:
+                # Check modification date first, fall back to creation date
+                date_str = metadata.get("modification_date") or metadata.get("creation_date")
+
+                if date_str:
+                    # Convert strings to datetime objects for proper comparison
+                    try:
+                        file_date = datetime.fromisoformat(date_str)
+                        start_dt = datetime.fromisoformat(start_time)
+                        end_dt = datetime.fromisoformat(end_time)
+
+                        # Compare dates properly
+                        if start_dt <= file_date <= end_dt:
+                            filtered_results.append(result)
+                    except (ValueError, TypeError):
+                        # Handle invalid date formats
+                        print(f"Warning: Invalid date format in metadata: {date_str}")
+                        continue
+
+        results = filtered_results
+
+    # Print results
+    print(f"\nSearch results for: '{query}'")
+    if time_matches:
+        print(
+            f"Time filter: {time_matches[0]['number']} {time_matches[0]['unit']}(s) {'(fuzzy)' if time_matches[0]['fuzzy'] else ''}"
+        )
+        print(
+            f"Date range: {time_matches[0]['range'][0][:10]} to {time_matches[0]['range'][1][:10]}"
+        )
+    print("-" * 80)
+
+    for i, result in enumerate(results, 1):
+        print(f"\n[{i}] Score: {result.score:.4f}")
+        print(f"Content: {result.text}")
+
+        # Show metadata if present
+        metadata = result.metadata if hasattr(result, "metadata") else None
+        if metadata:
+            if "creation_date" in metadata:
+                print(f"Created: {metadata['creation_date']}")
+            if "modification_date" in metadata:
+                print(f"Modified: {metadata['modification_date']}")
+        print("-" * 80)
+
+
+if __name__ == "__main__":
+    if len(sys.argv) < 2:
+        print('Usage: python search_index.py "<search query>" [top_k]')
+        sys.exit(1)
+
+    query = sys.argv[1]
+    top_k = int(sys.argv[2]) if len(sys.argv) > 2 else 15
+
+    search_files(query, top_k)

apps/semantic_file_search/leann_index_builder.py

@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+import json
+import sys
+from pathlib import Path
+
+from leann import LeannBuilder
+
+
+def process_json_items(json_file_path):
+    """Load and process JSON file with metadata items"""
+
+    with open(json_file_path, encoding="utf-8") as f:
+        items = json.load(f)
+
+    # Guard against empty JSON
+    if not items:
+        print("⚠️  No items found in the JSON file. Exiting gracefully.")
+        return
+
+    INDEX_PATH = str(Path("./").resolve() / "demo.leann")
+    builder = LeannBuilder(backend_name="hnsw", is_recompute=False)
+
+    total_items = len(items)
+    items_added = 0
+    print(f"Processing {total_items} items...")
+
+    for idx, item in enumerate(items):
+        try:
+            # Create embedding text sentence
+            embedding_text = f"{item.get('Name', 'unknown')} located at {item.get('Path', 'unknown')} and size {item.get('Size', 'unknown')} bytes with content type {item.get('ContentType', 'unknown')} and kind {item.get('Kind', 'unknown')}"
+
+            # Prepare metadata with dates
+            metadata = {}
+            if "CreationDate" in item:
+                metadata["creation_date"] = item["CreationDate"]
+            if "ContentChangeDate" in item:
+                metadata["modification_date"] = item["ContentChangeDate"]
+
+            # Add to builder
+            builder.add_text(embedding_text, metadata=metadata)
+            items_added += 1
+
+        except Exception as e:
+            print(f"\n⚠️  Warning: Failed to process item {idx}: {e}")
+            continue
+
+        # Show progress
+        progress = (idx + 1) / total_items * 100
+        sys.stdout.write(f"\rProgress: {idx + 1}/{total_items} ({progress:.1f}%)")
+        sys.stdout.flush()
+
+    print()  # New line after progress
+
+    # Guard against no successfully added items
+    if items_added == 0:
+        print("⚠️  No items were successfully added to the index. Exiting gracefully.")
+        return
+
+    print(f"\n✅ Successfully processed {items_added}/{total_items} items")
+    print("Building index...")
+
+    try:
+        builder.build_index(INDEX_PATH)
+        print(f"✓ Index saved to {INDEX_PATH}")
+    except ValueError as e:
+        if "No chunks added" in str(e):
+            print("⚠️  No chunks were added to the builder. Index not created.")
+        else:
+            raise
+
+
+if __name__ == "__main__":
+    if len(sys.argv) != 2:
+        print("Usage: python build_index.py <json_file>")
+        sys.exit(1)
+
+    json_file = sys.argv[1]
+    if not Path(json_file).exists():
+        print(f"Error: File {json_file} not found")
+        sys.exit(1)
+
+    process_json_items(json_file)

apps/semantic_file_search/spotlight_index_dump.py

@@ -0,0 +1,265 @@
+#!/usr/bin/env python3
+"""
+Spotlight Metadata Dumper for Vector DB
+Extracts only essential metadata for semantic search embeddings
+Output is optimized for vector database storage with minimal fields
+"""
+
+import json
+import sys
+from datetime import datetime
+
+# Check platform before importing macOS-specific modules
+if sys.platform != "darwin":
+    print("This script requires macOS (uses Spotlight)")
+    sys.exit(1)
+
+from Foundation import NSDate, NSMetadataQuery, NSPredicate, NSRunLoop
+
+# EDIT THIS LIST: Add or remove folders to search
+# Can be either:
+# - Folder names relative to home directory (e.g., "Desktop", "Downloads")
+# - Absolute paths (e.g., "/Applications", "/System/Library")
+SEARCH_FOLDERS = [
+    "Desktop",
+    "Downloads",
+    "Documents",
+    "Music",
+    "Pictures",
+    "Movies",
+    # "Library",  # Uncomment to include
+    # "/Applications",  # Absolute path example
+    # "Code/Projects",  # Subfolder example
+    # Add any other folders here
+]
+
+
+def convert_to_serializable(obj):
+    """Convert NS objects to Python serializable types"""
+    if obj is None:
+        return None
+
+    # Handle NSDate
+    if hasattr(obj, "timeIntervalSince1970"):
+        return datetime.fromtimestamp(obj.timeIntervalSince1970()).isoformat()
+
+    # Handle NSArray
+    if hasattr(obj, "count") and hasattr(obj, "objectAtIndex_"):
+        return [convert_to_serializable(obj.objectAtIndex_(i)) for i in range(obj.count())]
+
+    # Convert to string
+    try:
+        return str(obj)
+    except Exception:
+        return repr(obj)
+
+
+def dump_spotlight_data(max_items=10, output_file="spotlight_dump.json"):
+    """
+    Dump Spotlight data using public.item predicate
+    """
+    # Build full paths from SEARCH_FOLDERS
+    import os
+
+    home_dir = os.path.expanduser("~")
+    search_paths = []
+
+    print("Search locations:")
+    for folder in SEARCH_FOLDERS:
+        # Check if it's an absolute path or relative
+        if folder.startswith("/"):
+            full_path = folder
+        else:
+            full_path = os.path.join(home_dir, folder)
+
+        if os.path.exists(full_path):
+            search_paths.append(full_path)
+            print(f"  ✓ {full_path}")
+        else:
+            print(f"  ✗ {full_path} (not found)")
+
+    if not search_paths:
+        print("No valid search paths found!")
+        return []
+
+    print(f"\nDumping {max_items} items from Spotlight (public.item)...")
+
+    # Create query with public.item predicate
+    query = NSMetadataQuery.alloc().init()
+    predicate = NSPredicate.predicateWithFormat_("kMDItemContentTypeTree CONTAINS 'public.item'")
+    query.setPredicate_(predicate)
+
+    # Set search scopes to our specific folders
+    query.setSearchScopes_(search_paths)
+
+    print("Starting query...")
+    query.startQuery()
+
+    # Wait for gathering to complete
+    run_loop = NSRunLoop.currentRunLoop()
+    print("Gathering results...")
+
+    # Let it gather for a few seconds
+    for i in range(50):  # 5 seconds max
+        run_loop.runMode_beforeDate_(
+            "NSDefaultRunLoopMode", NSDate.dateWithTimeIntervalSinceNow_(0.1)
+        )
+        # Check gathering status periodically
+        if i % 10 == 0:
+            current_count = query.resultCount()
+            if current_count > 0:
+                print(f"  Found {current_count} items so far...")
+
+    # Continue while still gathering (up to 2 more seconds)
+    timeout = NSDate.dateWithTimeIntervalSinceNow_(2.0)
+    while query.isGathering() and timeout.timeIntervalSinceNow() > 0:
+        run_loop.runMode_beforeDate_(
+            "NSDefaultRunLoopMode", NSDate.dateWithTimeIntervalSinceNow_(0.1)
+        )
+
+    query.stopQuery()
+
+    total_results = query.resultCount()
+    print(f"Found {total_results} total items")
+
+    if total_results == 0:
+        print("No results found")
+        return []
+
+    # Process items
+    items_to_process = min(total_results, max_items)
+    results = []
+
+    # ONLY relevant attributes for vector embeddings
+    # These provide essential context for semantic search without bloat
+    attributes = [
+        "kMDItemPath",  # Full path for file retrieval
+        "kMDItemFSName",  # Filename for display & embedding
+        "kMDItemFSSize",  # Size for filtering/ranking
+        "kMDItemContentType",  # File type for categorization
+        "kMDItemKind",  # Human-readable type for embedding
+        "kMDItemFSCreationDate",  # Temporal context
+        "kMDItemFSContentChangeDate",  # Recency for ranking
+    ]
+
+    print(f"Processing {items_to_process} items...")
+
+    for i in range(items_to_process):
+        try:
+            item = query.resultAtIndex_(i)
+            metadata = {}
+
+            # Extract ONLY the relevant attributes
+            for attr in attributes:
+                try:
+                    value = item.valueForAttribute_(attr)
+                    if value is not None:
+                        # Keep the attribute name clean (remove kMDItem prefix for cleaner JSON)
+                        clean_key = attr.replace("kMDItem", "").replace("FS", "")
+                        metadata[clean_key] = convert_to_serializable(value)
+                except (AttributeError, ValueError, TypeError):
+                    continue
+
+            # Only add if we have at least a path
+            if metadata.get("Path"):
+                results.append(metadata)
+
+        except Exception as e:
+            print(f"Error processing item {i}: {e}")
+            continue
+
+    # Save to JSON
+    with open(output_file, "w", encoding="utf-8") as f:
+        json.dump(results, f, indent=2, ensure_ascii=False)
+
+    print(f"\n✓ Saved {len(results)} items to {output_file}")
+
+    # Show summary
+    print("\nSample items:")
+    import os
+
+    home_dir = os.path.expanduser("~")
+
+    for i, item in enumerate(results[:3]):
+        print(f"\n[Item {i + 1}]")
+        print(f"  Path: {item.get('Path', 'N/A')}")
+        print(f"  Name: {item.get('Name', 'N/A')}")
+        print(f"  Type: {item.get('ContentType', 'N/A')}")
+        print(f"  Kind: {item.get('Kind', 'N/A')}")
+
+        # Handle size properly
+        size = item.get("Size")
+        if size:
+            try:
+                size_int = int(size)
+                if size_int > 1024 * 1024:
+                    print(f"  Size: {size_int / (1024 * 1024):.2f} MB")
+                elif size_int > 1024:
+                    print(f"  Size: {size_int / 1024:.2f} KB")
+                else:
+                    print(f"  Size: {size_int} bytes")
+            except (ValueError, TypeError):
+                print(f"  Size: {size}")
+
+        # Show dates
+        if "CreationDate" in item:
+            print(f"  Created: {item['CreationDate']}")
+        if "ContentChangeDate" in item:
+            print(f"  Modified: {item['ContentChangeDate']}")
+
+    # Count by type
+    type_counts = {}
+    for item in results:
+        content_type = item.get("ContentType", "unknown")
+        type_counts[content_type] = type_counts.get(content_type, 0) + 1
+
+    print(f"\nTotal items saved: {len(results)}")
+
+    if type_counts:
+        print("\nTop content types:")
+        for ct, count in sorted(type_counts.items(), key=lambda x: x[1], reverse=True)[:5]:
+            print(f"  {ct}: {count} items")
+
+    # Count by folder
+    folder_counts = {}
+    for item in results:
+        path = item.get("Path", "")
+        for folder in SEARCH_FOLDERS:
+            # Build the full folder path
+            if folder.startswith("/"):
+                folder_path = folder
+            else:
+                folder_path = os.path.join(home_dir, folder)
+
+            if path.startswith(folder_path):
+                folder_counts[folder] = folder_counts.get(folder, 0) + 1
+                break
+
+    if folder_counts:
+        print("\nItems by location:")
+        for folder, count in sorted(folder_counts.items(), key=lambda x: x[1], reverse=True):
+            print(f"  {folder}: {count} items")
+
+    return results
+
+
+def main():
+    # Parse arguments
+    if len(sys.argv) > 1:
+        try:
+            max_items = int(sys.argv[1])
+        except ValueError:
+            print("Usage: python spot.py [number_of_items]")
+            print("Default: 10 items")
+            sys.exit(1)
+    else:
+        max_items = 10
+
+    output_file = sys.argv[2] if len(sys.argv) > 2 else "spotlight_dump.json"
+
+    # Run dump
+    dump_spotlight_data(max_items=max_items, output_file=output_file)
+
+
+if __name__ == "__main__":
+    main()

apps/wechat_rag.py

@@ -0,0 +1,189 @@
+"""
+WeChat History RAG example using the unified interface.
+Supports WeChat chat history export and search.
+"""
+
+import subprocess
+import sys
+from pathlib import Path
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+from base_rag_example import BaseRAGExample
+
+from .history_data.wechat_history import WeChatHistoryReader
+
+
+class WeChatRAG(BaseRAGExample):
+    """RAG example for WeChat chat history."""
+
+    def __init__(self):
+        # Set default values BEFORE calling super().__init__
+        self.max_items_default = -1  # Match original default
+        self.embedding_model_default = (
+            "sentence-transformers/all-MiniLM-L6-v2"  # Fast 384-dim model
+        )
+
+        super().__init__(
+            name="WeChat History",
+            description="Process and query WeChat chat history with LEANN",
+            default_index_name="wechat_history_magic_test_11Debug_new",
+        )
+
+    def _add_specific_arguments(self, parser):
+        """Add WeChat-specific arguments."""
+        wechat_group = parser.add_argument_group("WeChat Parameters")
+        wechat_group.add_argument(
+            "--export-dir",
+            type=str,
+            default="./wechat_export",
+            help="Directory to store WeChat exports (default: ./wechat_export)",
+        )
+        wechat_group.add_argument(
+            "--force-export",
+            action="store_true",
+            help="Force re-export of WeChat data even if exports exist",
+        )
+        wechat_group.add_argument(
+            "--chunk-size", type=int, default=192, help="Text chunk size (default: 192)"
+        )
+        wechat_group.add_argument(
+            "--chunk-overlap", type=int, default=64, help="Text chunk overlap (default: 64)"
+        )
+
+    def _export_wechat_data(self, export_dir: Path) -> bool:
+        """Export WeChat data using wechattweak-cli."""
+        print("Exporting WeChat data...")
+
+        # Check if WeChat is running
+        try:
+            result = subprocess.run(["pgrep", "WeChat"], capture_output=True, text=True)
+            if result.returncode != 0:
+                print("WeChat is not running. Please start WeChat first.")
+                return False
+        except Exception:
+            pass  # pgrep might not be available on all systems
+
+        # Create export directory
+        export_dir.mkdir(parents=True, exist_ok=True)
+
+        # Run export command
+        cmd = ["packages/wechat-exporter/wechattweak-cli", "export", str(export_dir)]
+
+        try:
+            print(f"Running: {' '.join(cmd)}")
+            result = subprocess.run(cmd, capture_output=True, text=True)
+
+            if result.returncode == 0:
+                print("WeChat data exported successfully!")
+                return True
+            else:
+                print(f"Export failed: {result.stderr}")
+                return False
+
+        except FileNotFoundError:
+            print("\nError: wechattweak-cli not found!")
+            print("Please install it first:")
+            print("  sudo packages/wechat-exporter/wechattweak-cli install")
+            return False
+        except Exception as e:
+            print(f"Export error: {e}")
+            return False
+
+    async def load_data(self, args) -> list[str]:
+        """Load WeChat history and convert to text chunks."""
+        # Initialize WeChat reader with export capabilities
+        reader = WeChatHistoryReader()
+
+        # Find existing exports or create new ones using the centralized method
+        export_dirs = reader.find_or_export_wechat_data(args.export_dir)
+        if not export_dirs:
+            print("Failed to find or export WeChat data. Trying to find any existing exports...")
+            # Try to find any existing exports in common locations
+            export_dirs = reader.find_wechat_export_dirs()
+            if not export_dirs:
+                print("No WeChat data found. Please ensure WeChat exports exist.")
+                return []
+
+        # Load documents from all found export directories
+        all_documents = []
+        total_processed = 0
+
+        for i, export_dir in enumerate(export_dirs):
+            print(f"\nProcessing WeChat export {i + 1}/{len(export_dirs)}: {export_dir}")
+
+            try:
+                # Apply max_items limit per export
+                max_per_export = -1
+                if args.max_items > 0:
+                    remaining = args.max_items - total_processed
+                    if remaining <= 0:
+                        break
+                    max_per_export = remaining
+
+                documents = reader.load_data(
+                    wechat_export_dir=str(export_dir),
+                    max_count=max_per_export,
+                    concatenate_messages=True,  # Enable message concatenation for better context
+                )
+
+                if documents:
+                    print(f"Loaded {len(documents)} chat documents from {export_dir}")
+                    all_documents.extend(documents)
+                    total_processed += len(documents)
+                else:
+                    print(f"No documents loaded from {export_dir}")
+
+            except Exception as e:
+                print(f"Error processing {export_dir}: {e}")
+                continue
+
+        if not all_documents:
+            print("No documents loaded from any source. Exiting.")
+            return []
+
+        print(f"\nTotal loaded {len(all_documents)} chat documents from {len(export_dirs)} exports")
+        print("now starting to split into text chunks ... take some time")
+
+        # Convert to text chunks with contact information
+        all_texts = []
+        for doc in all_documents:
+            # Split the document into chunks
+            from llama_index.core.node_parser import SentenceSplitter
+
+            text_splitter = SentenceSplitter(
+                chunk_size=args.chunk_size, chunk_overlap=args.chunk_overlap
+            )
+            nodes = text_splitter.get_nodes_from_documents([doc])
+
+            for node in nodes:
+                # Add contact information to each chunk
+                contact_name = doc.metadata.get("contact_name", "Unknown")
+                text = f"[Contact] means the message is from: {contact_name}\n" + node.get_content()
+                all_texts.append(text)
+
+        print(f"Created {len(all_texts)} text chunks from {len(all_documents)} documents")
+        return all_texts
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Check platform
+    if sys.platform != "darwin":
+        print("\n⚠️  Warning: WeChat export is only supported on macOS")
+        print("   You can still query existing exports on other platforms\n")
+
+    # Example queries for WeChat RAG
+    print("\n💬 WeChat History RAG Example")
+    print("=" * 50)
+    print("\nExample queries you can try:")
+    print("- 'Show me conversations about travel plans'")
+    print("- 'Find group chats about weekend activities'")
+    print("- '我想买魔术师约翰逊的球衣,给我一些对应聊天记录?'")
+    print("- 'What did we discuss about the project last month?'")
+    print("\nNote: WeChat must be running for export to work\n")
+
+    rag = WeChatRAG()
+    asyncio.run(rag.run())

benchmarks/README.md

@@ -1,13 +1,28 @@
-# 🧪 Leann Sanity Checks
+# 🧪 LEANN Benchmarks & Testing
 
-This directory contains comprehensive sanity checks for the Leann system, ensuring all components work correctly across different configurations.
+This directory contains performance benchmarks and comprehensive tests for the LEANN system, including backend comparisons and sanity checks across different configurations.
 
 ## 📁 Test Files
 
+### `diskann_vs_hnsw_speed_comparison.py`
+Performance comparison between DiskANN and HNSW backends:
+- ✅ **Search latency** comparison with both backends using recompute
+- ✅ **Index size** and **build time** measurements
+- ✅ **Score validity** testing (ensures no -inf scores)
+- ✅ **Configurable dataset sizes** for different scales
+
+```bash
+# Quick comparison with 500 docs, 10 queries
+python benchmarks/diskann_vs_hnsw_speed_comparison.py
+
+# Large-scale comparison with 2000 docs, 20 queries
+python benchmarks/diskann_vs_hnsw_speed_comparison.py 2000 20
+```
+
 ### `test_distance_functions.py`
 Tests all supported distance functions across DiskANN backend:
 - ✅ **MIPS** (Maximum Inner Product Search)
-- ✅ **L2** (Euclidean Distance) 
+- ✅ **L2** (Euclidean Distance)
 - ✅ **Cosine** (Cosine Similarity)
 
 ```bash
@@ -27,7 +42,7 @@ uv run python tests/sanity_checks/test_l2_verification.py
 ### `test_sanity_check.py`
 Comprehensive end-to-end verification including:
 - Distance function testing
-- Embedding model compatibility  
+- Embedding model compatibility
 - Search result correctness validation
 - Backend integration testing
 
@@ -64,7 +79,7 @@ When all tests pass, you should see:
 ```
 📊 测试结果总结:
   mips      : ✅ 通过
-  l2        : ✅ 通过  
+  l2        : ✅ 通过
   cosine    : ✅ 通过
 
 🎉 测试完成!
@@ -98,7 +113,7 @@ pkill -f "embedding_server"
 
 ### Typical Timing (3 documents, consumer hardware):
 - **Index Building**: 2-5 seconds per distance function
-- **Search Query**: 50-200ms 
+- **Search Query**: 50-200ms
 - **Recompute Mode**: 5-15 seconds (higher accuracy)
 
 ### Memory Usage:
@@ -117,4 +132,4 @@ These tests are designed to be run in automated environments:
     uv run python tests/sanity_checks/test_l2_verification.py
 ```
 
-The tests are deterministic and should produce consistent results across different platforms.
+The tests are deterministic and should produce consistent results across different platforms.

benchmarks/benchmark_embeddings.py

@@ -1,43 +1,46 @@
 import time
-import numpy as np
+
 import matplotlib.pyplot as plt
-import torch
-from sentence_transformers import SentenceTransformer
 import mlx.core as mx
+import numpy as np
+import torch
 from mlx_lm import load
+from sentence_transformers import SentenceTransformer
 
 # --- Configuration ---
 MODEL_NAME_TORCH = "Qwen/Qwen3-Embedding-0.6B"
 MODEL_NAME_MLX = "mlx-community/Qwen3-Embedding-0.6B-4bit-DWQ"
 BATCH_SIZES = [1, 8, 16, 32, 64, 128]
 NUM_RUNS = 10  # Number of runs to average for each batch size
-WARMUP_RUNS = 2 # Number of warm-up runs
+WARMUP_RUNS = 2  # Number of warm-up runs
 
 # --- Generate Dummy Data ---
 DUMMY_SENTENCES = ["This is a test sentence for benchmarking." * 5] * max(BATCH_SIZES)
 
 # --- Benchmark Functions ---b
 
+
 def benchmark_torch(model, sentences):
     start_time = time.time()
     model.encode(sentences, convert_to_numpy=True)
     end_time = time.time()
     return (end_time - start_time) * 1000  # Return time in ms
 
+
 def benchmark_mlx(model, tokenizer, sentences):
     start_time = time.time()
-    
+
     # Tokenize sentences using MLX tokenizer
     tokens = []
     for sentence in sentences:
         token_ids = tokenizer.encode(sentence)
         tokens.append(token_ids)
-    
+
     # Pad sequences to the same length
     max_len = max(len(t) for t in tokens)
     input_ids = []
     attention_mask = []
-    
+
     for token_seq in tokens:
         # Pad sequence
         padded = token_seq + [tokenizer.eos_token_id] * (max_len - len(token_seq))
@@ -45,24 +48,25 @@ def benchmark_mlx(model, tokenizer, sentences):
         # Create attention mask (1 for real tokens, 0 for padding)
         mask = [1] * len(token_seq) + [0] * (max_len - len(token_seq))
         attention_mask.append(mask)
-    
+
     # Convert to MLX arrays
     input_ids = mx.array(input_ids)
     attention_mask = mx.array(attention_mask)
-    
+
     # Get embeddings
     embeddings = model(input_ids)
-    
+
     # Mean pooling
     mask = mx.expand_dims(attention_mask, -1)
     sum_embeddings = (embeddings * mask).sum(axis=1)
     sum_mask = mask.sum(axis=1)
     _ = sum_embeddings / sum_mask
-    
+
     mx.eval()  # Ensure computation is finished
     end_time = time.time()
     return (end_time - start_time) * 1000  # Return time in ms
 
+
 # --- Main Execution ---
 def main():
     print("--- Initializing Models ---")
@@ -92,13 +96,15 @@ def main():
     for batch_size in BATCH_SIZES:
         print(f"Benchmarking batch size: {batch_size}")
         sentences_batch = DUMMY_SENTENCES[:batch_size]
-        
+
         # Benchmark PyTorch
         torch_times = [benchmark_torch(model_torch, sentences_batch) for _ in range(NUM_RUNS)]
         results_torch.append(np.mean(torch_times))
-        
+
         # Benchmark MLX
-        mlx_times = [benchmark_mlx(model_mlx, tokenizer_mlx, sentences_batch) for _ in range(NUM_RUNS)]
+        mlx_times = [
+            benchmark_mlx(model_mlx, tokenizer_mlx, sentences_batch) for _ in range(NUM_RUNS)
+        ]
         results_mlx.append(np.mean(mlx_times))
 
     print("\n--- Benchmark Results (Average time per batch in ms) ---")
@@ -109,20 +115,27 @@ def main():
     # --- Plotting ---
     print("\n--- Generating Plot ---")
     plt.figure(figsize=(10, 6))
-    plt.plot(BATCH_SIZES, results_torch, marker='o', linestyle='-', label=f'PyTorch ({device})')
-    plt.plot(BATCH_SIZES, results_mlx, marker='s', linestyle='-', label='MLX')
+    plt.plot(
+        BATCH_SIZES,
+        results_torch,
+        marker="o",
+        linestyle="-",
+        label=f"PyTorch ({device})",
+    )
+    plt.plot(BATCH_SIZES, results_mlx, marker="s", linestyle="-", label="MLX")
 
-    plt.title(f'Embedding Performance: MLX vs PyTorch\nModel: {MODEL_NAME_TORCH}')
+    plt.title(f"Embedding Performance: MLX vs PyTorch\nModel: {MODEL_NAME_TORCH}")
     plt.xlabel("Batch Size")
     plt.ylabel("Average Time per Batch (ms)")
     plt.xticks(BATCH_SIZES)
     plt.grid(True)
     plt.legend()
-    
+
     # Save the plot
     output_filename = "embedding_benchmark.png"
     plt.savefig(output_filename)
     print(f"Plot saved to {output_filename}")
 
+
 if __name__ == "__main__":
     main()

benchmarks/benchmark_no_recompute.py

@@ -0,0 +1,148 @@
+import argparse
+import os
+import time
+from pathlib import Path
+
+from leann import LeannBuilder, LeannSearcher
+
+
+def _meta_exists(index_path: str) -> bool:
+    p = Path(index_path)
+    return (p.parent / f"{p.stem}.meta.json").exists()
+
+
+def ensure_index(index_path: str, backend_name: str, num_docs: int, is_recompute: bool) -> None:
+    # if _meta_exists(index_path):
+    #     return
+    kwargs = {}
+    if backend_name == "hnsw":
+        kwargs["is_compact"] = is_recompute
+    builder = LeannBuilder(
+        backend_name=backend_name,
+        embedding_model=os.getenv("LEANN_EMBED_MODEL", "facebook/contriever"),
+        embedding_mode=os.getenv("LEANN_EMBED_MODE", "sentence-transformers"),
+        graph_degree=32,
+        complexity=64,
+        is_recompute=is_recompute,
+        num_threads=4,
+        **kwargs,
+    )
+    for i in range(num_docs):
+        builder.add_text(
+            f"This is a test document number {i}. It contains some repeated text for benchmarking."
+        )
+    builder.build_index(index_path)
+
+
+def _bench_group(
+    index_path: str,
+    recompute: bool,
+    query: str,
+    repeats: int,
+    complexity: int = 32,
+    top_k: int = 10,
+) -> float:
+    # Independent searcher per group; fixed port when recompute
+    searcher = LeannSearcher(index_path=index_path)
+
+    # Warm-up once
+    _ = searcher.search(
+        query,
+        top_k=top_k,
+        complexity=complexity,
+        recompute_embeddings=recompute,
+    )
+
+    def _once() -> float:
+        t0 = time.time()
+        _ = searcher.search(
+            query,
+            top_k=top_k,
+            complexity=complexity,
+            recompute_embeddings=recompute,
+        )
+        return time.time() - t0
+
+    if repeats <= 1:
+        t = _once()
+    else:
+        vals = [_once() for _ in range(repeats)]
+        vals.sort()
+        t = vals[len(vals) // 2]
+
+    searcher.cleanup()
+    return t
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--num-docs", type=int, default=5000)
+    parser.add_argument("--repeats", type=int, default=3)
+    parser.add_argument("--complexity", type=int, default=32)
+    args = parser.parse_args()
+
+    base = Path.cwd() / ".leann" / "indexes" / f"bench_n{args.num_docs}"
+    base.parent.mkdir(parents=True, exist_ok=True)
+    # ---------- Build HNSW variants ----------
+    hnsw_r = str(base / f"hnsw_recompute_n{args.num_docs}.leann")
+    hnsw_nr = str(base / f"hnsw_norecompute_n{args.num_docs}.leann")
+    ensure_index(hnsw_r, "hnsw", args.num_docs, True)
+    ensure_index(hnsw_nr, "hnsw", args.num_docs, False)
+
+    # ---------- Build DiskANN variants ----------
+    diskann_r = str(base / "diskann_r.leann")
+    diskann_nr = str(base / "diskann_nr.leann")
+    ensure_index(diskann_r, "diskann", args.num_docs, True)
+    ensure_index(diskann_nr, "diskann", args.num_docs, False)
+
+    # ---------- Helpers ----------
+    def _size_for(prefix: str) -> int:
+        p = Path(prefix)
+        base_dir = p.parent
+        stem = p.stem
+        total = 0
+        for f in base_dir.iterdir():
+            if f.is_file() and f.name.startswith(stem):
+                total += f.stat().st_size
+        return total
+
+    # ---------- HNSW benchmark ----------
+    t_hnsw_r = _bench_group(
+        hnsw_r, True, "test document number 42", repeats=args.repeats, complexity=args.complexity
+    )
+    t_hnsw_nr = _bench_group(
+        hnsw_nr, False, "test document number 42", repeats=args.repeats, complexity=args.complexity
+    )
+    size_hnsw_r = _size_for(hnsw_r)
+    size_hnsw_nr = _size_for(hnsw_nr)
+
+    print("Benchmark results (HNSW):")
+    print(f"  recompute=True:  search_time={t_hnsw_r:.3f}s, size={size_hnsw_r / 1024 / 1024:.1f}MB")
+    print(
+        f"  recompute=False: search_time={t_hnsw_nr:.3f}s, size={size_hnsw_nr / 1024 / 1024:.1f}MB"
+    )
+    print("  Expectation: no-recompute should be faster but larger on disk.")
+
+    # ---------- DiskANN benchmark ----------
+    t_diskann_r = _bench_group(
+        diskann_r, True, "DiskANN R test doc 123", repeats=args.repeats, complexity=args.complexity
+    )
+    t_diskann_nr = _bench_group(
+        diskann_nr,
+        False,
+        "DiskANN NR test doc 123",
+        repeats=args.repeats,
+        complexity=args.complexity,
+    )
+    size_diskann_r = _size_for(diskann_r)
+    size_diskann_nr = _size_for(diskann_nr)
+
+    print("\nBenchmark results (DiskANN):")
+    print(f"  build(recompute=True, partition): size={size_diskann_r / 1024 / 1024:.1f}MB")
+    print(f"  build(recompute=False):          size={size_diskann_nr / 1024 / 1024:.1f}MB")
+    print(f"  search recompute=True (final rerank): {t_diskann_r:.3f}s")
+    print(f"  search recompute=False (PQ only):     {t_diskann_nr:.3f}s")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/bm25_diskann_baselines/README.md

@@ -0,0 +1,23 @@
+BM25 vs DiskANN Baselines
+
+```bash
+aws s3 sync s3://powerrag-diskann-rpj-wiki-20250824-224037-194d640c/bm25_rpj_wiki/index_en_only/ benchmarks/data/indices/bm25_index/
+aws s3 sync s3://powerrag-diskann-rpj-wiki-20250824-224037-194d640c/diskann_rpj_wiki/ benchmarks/data/indices/diskann_rpj_wiki/
+```
+
+- Dataset: `benchmarks/data/queries/nq_open.jsonl` (Natural Questions)
+- Machine-specific; results measured locally with the current repo.
+
+DiskANN (NQ queries, search-only)
+- Command: `uv run --script benchmarks/bm25_diskann_baselines/run_diskann.py`
+- Settings: `recompute_embeddings=False`, embeddings precomputed (excluded from timing), batching off, caching off (`cache_mechanism=2`, `num_nodes_to_cache=0`)
+- Result: avg 0.011093 s/query, QPS 90.15 (p50 0.010731 s, p95 0.015000 s)
+
+BM25
+- Command: `uv run --script benchmarks/bm25_diskann_baselines/run_bm25.py`
+- Settings: `k=10`, `k1=0.9`, `b=0.4`, queries=100
+- Result: avg 0.028589 s/query, QPS 34.97 (p50 0.026060 s, p90 0.043695 s, p95 0.053260 s, p99 0.055257 s)
+
+Notes
+- DiskANN measures search-only latency on real NQ queries (embeddings computed beforehand and excluded from timing).
+- Use `benchmarks/bm25_diskann_baselines/run_diskann.py` for DiskANN; `benchmarks/bm25_diskann_baselines/run_bm25.py` for BM25.

benchmarks/bm25_diskann_baselines/run_bm25.py

@@ -0,0 +1,183 @@
+# /// script
+# dependencies = [
+#   "pyserini"
+# ]
+# ///
+# sudo pacman -S jdk21-openjdk
+# export JAVA_HOME=/usr/lib/jvm/java-21-openjdk
+# sudo archlinux-java status
+# sudo archlinux-java set java-21-openjdk
+# set -Ux JAVA_HOME /usr/lib/jvm/java-21-openjdk
+# fish_add_path --global $JAVA_HOME/bin
+# set -Ux LD_LIBRARY_PATH $JAVA_HOME/lib/server $LD_LIBRARY_PATH
+# which javac # Should be /usr/lib/jvm/java-21-openjdk/bin/javac
+
+import argparse
+import json
+import os
+import sys
+import time
+from statistics import mean
+
+
+def load_queries(path: str, limit: int | None) -> list[str]:
+    queries: list[str] = []
+    # Try JSONL with a 'query' or 'text' field; fallback to plain text (one query per line)
+    _, ext = os.path.splitext(path)
+    if ext.lower() in {".jsonl", ".json"}:
+        with open(path, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    obj = json.loads(line)
+                except json.JSONDecodeError:
+                    # Not strict JSONL? treat the whole line as the query
+                    queries.append(line)
+                    continue
+                q = obj.get("query") or obj.get("text") or obj.get("question")
+                if q:
+                    queries.append(str(q))
+    else:
+        with open(path, encoding="utf-8") as f:
+            for line in f:
+                s = line.strip()
+                if s:
+                    queries.append(s)
+
+    if limit is not None and limit > 0:
+        queries = queries[:limit]
+    return queries
+
+
+def percentile(values: list[float], p: float) -> float:
+    if not values:
+        return 0.0
+    s = sorted(values)
+    k = (len(s) - 1) * (p / 100.0)
+    f = int(k)
+    c = min(f + 1, len(s) - 1)
+    if f == c:
+        return s[f]
+    return s[f] + (s[c] - s[f]) * (k - f)
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Standalone BM25 latency benchmark (Pyserini)")
+    ap.add_argument(
+        "--bm25-index",
+        default="benchmarks/data/indices/bm25_index",
+        help="Path to Pyserini Lucene index directory",
+    )
+    ap.add_argument(
+        "--queries",
+        default="benchmarks/data/queries/nq_open.jsonl",
+        help="Path to queries file (JSONL with 'query'/'text' or plain txt one-per-line)",
+    )
+    ap.add_argument("--k", type=int, default=10, help="Top-k to retrieve (default: 10)")
+    ap.add_argument("--k1", type=float, default=0.9, help="BM25 k1 (default: 0.9)")
+    ap.add_argument("--b", type=float, default=0.4, help="BM25 b (default: 0.4)")
+    ap.add_argument("--limit", type=int, default=100, help="Max queries to run (default: 100)")
+    ap.add_argument(
+        "--warmup", type=int, default=5, help="Warmup queries not counted in latency (default: 5)"
+    )
+    ap.add_argument(
+        "--fetch-docs", action="store_true", help="Also fetch doc contents (slower; default: off)"
+    )
+    ap.add_argument("--report", type=str, default=None, help="Optional JSON report path")
+    args = ap.parse_args()
+
+    try:
+        from pyserini.search.lucene import LuceneSearcher
+    except Exception:
+        print("Pyserini not found. Install with: pip install pyserini", file=sys.stderr)
+        raise
+
+    if not os.path.isdir(args.bm25_index):
+        print(f"Index directory not found: {args.bm25_index}", file=sys.stderr)
+        sys.exit(1)
+
+    queries = load_queries(args.queries, args.limit)
+    if not queries:
+        print("No queries loaded.", file=sys.stderr)
+        sys.exit(1)
+
+    print(f"Loaded {len(queries)} queries from {args.queries}")
+    print(f"Opening BM25 index: {args.bm25_index}")
+    searcher = LuceneSearcher(args.bm25_index)
+    # Some builds of pyserini require explicit set_bm25; others ignore
+    try:
+        searcher.set_bm25(k1=args.k1, b=args.b)
+    except Exception:
+        pass
+
+    latencies: list[float] = []
+    total_searches = 0
+
+    # Warmup
+    for i in range(min(args.warmup, len(queries))):
+        _ = searcher.search(queries[i], k=args.k)
+
+    t0 = time.time()
+    for i, q in enumerate(queries):
+        t1 = time.time()
+        hits = searcher.search(q, k=args.k)
+        t2 = time.time()
+        latencies.append(t2 - t1)
+        total_searches += 1
+
+        if args.fetch_docs:
+            # Optional doc fetch to include I/O time
+            for h in hits:
+                try:
+                    _ = searcher.doc(h.docid)
+                except Exception:
+                    pass
+
+        if (i + 1) % 50 == 0:
+            print(f"Processed {i + 1}/{len(queries)} queries")
+
+    t1 = time.time()
+    total_time = t1 - t0
+
+    if latencies:
+        avg = mean(latencies)
+        p50 = percentile(latencies, 50)
+        p90 = percentile(latencies, 90)
+        p95 = percentile(latencies, 95)
+        p99 = percentile(latencies, 99)
+        qps = total_searches / total_time if total_time > 0 else 0.0
+    else:
+        avg = p50 = p90 = p95 = p99 = qps = 0.0
+
+    print("BM25 Latency Report")
+    print(f"  queries: {total_searches}")
+    print(f"  k: {args.k}, k1: {args.k1}, b: {args.b}")
+    print(f"  avg per query: {avg:.6f} s")
+    print(f"  p50/p90/p95/p99: {p50:.6f}/{p90:.6f}/{p95:.6f}/{p99:.6f} s")
+    print(f"  total time: {total_time:.3f} s, qps: {qps:.2f}")
+
+    if args.report:
+        payload = {
+            "queries": total_searches,
+            "k": args.k,
+            "k1": args.k1,
+            "b": args.b,
+            "avg_s": avg,
+            "p50_s": p50,
+            "p90_s": p90,
+            "p95_s": p95,
+            "p99_s": p99,
+            "total_time_s": total_time,
+            "qps": qps,
+            "index_dir": os.path.abspath(args.bm25_index),
+            "fetch_docs": bool(args.fetch_docs),
+        }
+        with open(args.report, "w", encoding="utf-8") as f:
+            json.dump(payload, f, indent=2)
+        print(f"Saved report to {args.report}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/bm25_diskann_baselines/run_diskann.py

@@ -0,0 +1,124 @@
+# /// script
+# dependencies = [
+#   "leann-backend-diskann"
+# ]
+# ///
+
+import argparse
+import json
+import time
+from pathlib import Path
+
+import numpy as np
+
+
+def load_queries(path: Path, limit: int | None) -> list[str]:
+    out: list[str] = []
+    with open(path, encoding="utf-8") as f:
+        for line in f:
+            obj = json.loads(line)
+            out.append(obj["query"])
+            if limit and len(out) >= limit:
+                break
+    return out
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(
+        description="DiskANN baseline on real NQ queries (search-only timing)"
+    )
+    ap.add_argument(
+        "--index-dir",
+        default="benchmarks/data/indices/diskann_rpj_wiki",
+        help="Directory containing DiskANN files",
+    )
+    ap.add_argument("--index-prefix", default="ann")
+    ap.add_argument("--queries-file", default="benchmarks/data/queries/nq_open.jsonl")
+    ap.add_argument("--num-queries", type=int, default=200)
+    ap.add_argument("--top-k", type=int, default=10)
+    ap.add_argument("--complexity", type=int, default=62)
+    ap.add_argument("--threads", type=int, default=1)
+    ap.add_argument("--beam-width", type=int, default=1)
+    ap.add_argument("--cache-mechanism", type=int, default=2)
+    ap.add_argument("--num-nodes-to-cache", type=int, default=0)
+    args = ap.parse_args()
+
+    index_dir = Path(args.index_dir).resolve()
+    if not index_dir.is_dir():
+        raise SystemExit(f"Index dir not found: {index_dir}")
+
+    qpath = Path(args.queries_file).resolve()
+    if not qpath.exists():
+        raise SystemExit(f"Queries file not found: {qpath}")
+
+    queries = load_queries(qpath, args.num_queries)
+    print(f"Loaded {len(queries)} queries from {qpath}")
+
+    # Compute embeddings once (exclude from timing)
+    from leann.api import compute_embeddings as _compute
+
+    embs = _compute(
+        queries,
+        model_name="facebook/contriever-msmarco",
+        mode="sentence-transformers",
+        use_server=False,
+    ).astype(np.float32)
+    if embs.ndim != 2:
+        raise SystemExit("Embedding compute failed or returned wrong shape")
+
+    # Build searcher
+    from leann_backend_diskann.diskann_backend import DiskannSearcher as _DiskannSearcher
+
+    index_prefix_path = str(index_dir / args.index_prefix)
+    searcher = _DiskannSearcher(
+        index_prefix_path,
+        num_threads=int(args.threads),
+        cache_mechanism=int(args.cache_mechanism),
+        num_nodes_to_cache=int(args.num_nodes_to_cache),
+    )
+
+    # Warmup (not timed)
+    _ = searcher.search(
+        embs[0:1],
+        top_k=args.top_k,
+        complexity=args.complexity,
+        beam_width=args.beam_width,
+        prune_ratio=0.0,
+        recompute_embeddings=False,
+        batch_recompute=False,
+        dedup_node_dis=False,
+    )
+
+    # Timed loop
+    times: list[float] = []
+    for i in range(embs.shape[0]):
+        t0 = time.time()
+        _ = searcher.search(
+            embs[i : i + 1],
+            top_k=args.top_k,
+            complexity=args.complexity,
+            beam_width=args.beam_width,
+            prune_ratio=0.0,
+            recompute_embeddings=False,
+            batch_recompute=False,
+            dedup_node_dis=False,
+        )
+        times.append(time.time() - t0)
+
+    times_sorted = sorted(times)
+    avg = float(sum(times) / len(times))
+    p50 = times_sorted[len(times) // 2]
+    p95 = times_sorted[max(0, int(len(times) * 0.95) - 1)]
+
+    print("\nDiskANN (NQ, search-only) Report")
+    print(f"  queries: {len(times)}")
+    print(
+        f"  k: {args.top_k}, complexity: {args.complexity}, beam_width: {args.beam_width}, threads: {args.threads}"
+    )
+    print(f"  avg per query: {avg:.6f} s")
+    print(f"  p50/p95: {p50:.6f}/{p95:.6f} s")
+    print(f"  QPS: {1.0 / avg:.2f}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/compare_faiss_vs_leann.py

@@ -3,14 +3,15 @@
 Memory comparison between Faiss HNSW and LEANN HNSW backend
 """
 
+import gc
 import logging
 import os
+import subprocess
 import sys
 import time
-import psutil
-import gc
-import subprocess
 from pathlib import Path
+
+import psutil
 from llama_index.core.node_parser import SentenceSplitter
 
 # Setup logging
@@ -61,7 +62,7 @@ def test_faiss_hnsw():
 
     try:
         result = subprocess.run(
-            [sys.executable, "examples/faiss_only.py"],
+            [sys.executable, "benchmarks/faiss_only.py"],
             capture_output=True,
             text=True,
             timeout=300,
@@ -83,9 +84,7 @@ def test_faiss_hnsw():
 
         for line in lines:
             if "Peak Memory:" in line:
-                peak_memory = float(
-                    line.split("Peak Memory:")[1].split("MB")[0].strip()
-                )
+                peak_memory = float(line.split("Peak Memory:")[1].split("MB")[0].strip())
 
         return {"peak_memory": peak_memory}
 
@@ -111,13 +110,12 @@ def test_leann_hnsw():
 
     tracker.checkpoint("After imports")
 
+    from leann.api import LeannBuilder
     from llama_index.core import SimpleDirectoryReader
-    from leann.api import LeannBuilder, LeannSearcher
-
 
     # Load and parse documents
     documents = SimpleDirectoryReader(
-        "examples/data",
+        "data",
         recursive=True,
         encoding="utf-8",
         required_exts=[".pdf", ".txt", ".md"],
@@ -135,6 +133,7 @@ def test_leann_hnsw():
         nodes = node_parser.get_nodes_from_documents([doc])
         for node in nodes:
             all_texts.append(node.get_content())
+    print(f"Total number of chunks: {len(all_texts)}")
 
     tracker.checkpoint("After text chunking")
 
@@ -196,16 +195,14 @@ def test_leann_hnsw():
     runtime_start_mem = get_memory_usage()
     print(f"Before load memory: {runtime_start_mem:.1f} MB")
     tracker.checkpoint("Before load memory")
-    
+
     # Load searcher
     searcher = LeannSearcher(index_path)
     tracker.checkpoint("After searcher loading")
 
-
-
     print("Running search queries...")
     queries = [
-        "什么是盘古大模型以及盘古开发过程中遇到了什么阴暗面，任务令一般在什么城市颁发",
+        "什么是盘古大模型以及盘古开发过程中遇到了什么阴暗面,任务令一般在什么城市颁发",
         "What is LEANN and how does it work?",
         "华为诺亚方舟实验室的主要研究内容",
     ]
@@ -303,21 +300,15 @@ def main():
 
         print("\nLEANN vs Faiss Performance:")
         memory_saving = faiss_results["peak_memory"] - leann_results["peak_memory"]
-        print(
-            f"  Search Memory: {memory_ratio:.1f}x less ({memory_saving:.1f} MB saved)"
-        )
+        print(f"  Search Memory: {memory_ratio:.1f}x less ({memory_saving:.1f} MB saved)")
 
         # Storage comparison
         if leann_storage_size > faiss_storage_size:
             storage_ratio = leann_storage_size / faiss_storage_size
-            print(
-                f"  Storage Size: {storage_ratio:.1f}x larger (LEANN uses more storage)"
-            )
+            print(f"  Storage Size: {storage_ratio:.1f}x larger (LEANN uses more storage)")
         elif faiss_storage_size > leann_storage_size:
             storage_ratio = faiss_storage_size / leann_storage_size
-            print(
-                f"  Storage Size: {storage_ratio:.1f}x smaller (LEANN uses less storage)"
-            )
+            print(f"  Storage Size: {storage_ratio:.1f}x smaller (LEANN uses less storage)")
         else:
             print("  Storage Size: similar")
     else:

benchmarks/diskann_vs_hnsw_speed_comparison.py

@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+"""
+DiskANN vs HNSW Search Performance Comparison
+
+This benchmark compares search performance between DiskANN and HNSW backends:
+- DiskANN: With graph partitioning enabled (is_recompute=True)
+- HNSW: With recompute enabled (is_recompute=True)
+- Tests performance across different dataset sizes
+- Measures search latency, recall, and index size
+"""
+
+import gc
+import multiprocessing as mp
+import tempfile
+import time
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+# Prefer 'fork' start method to avoid POSIX semaphore leaks on macOS
+try:
+    mp.set_start_method("fork", force=True)
+except Exception:
+    pass
+
+
+def create_test_texts(n_docs: int) -> list[str]:
+    """Create synthetic test documents for benchmarking."""
+    np.random.seed(42)
+    topics = [
+        "machine learning and artificial intelligence",
+        "natural language processing and text analysis",
+        "computer vision and image recognition",
+        "data science and statistical analysis",
+        "deep learning and neural networks",
+        "information retrieval and search engines",
+        "database systems and data management",
+        "software engineering and programming",
+        "cybersecurity and network protection",
+        "cloud computing and distributed systems",
+    ]
+
+    texts = []
+    for i in range(n_docs):
+        topic = topics[i % len(topics)]
+        variation = np.random.randint(1, 100)
+        text = (
+            f"This is document {i} about {topic}. Content variation {variation}. "
+            f"Additional information about {topic} with details and examples. "
+            f"Technical discussion of {topic} including implementation aspects."
+        )
+        texts.append(text)
+
+    return texts
+
+
+def benchmark_backend(
+    backend_name: str, texts: list[str], test_queries: list[str], backend_kwargs: dict[str, Any]
+) -> dict[str, float]:
+    """Benchmark a specific backend with the given configuration."""
+    from leann.api import LeannBuilder, LeannSearcher
+
+    print(f"\n🔧 Testing {backend_name.upper()} backend...")
+
+    with tempfile.TemporaryDirectory() as temp_dir:
+        index_path = str(Path(temp_dir) / f"benchmark_{backend_name}.leann")
+
+        # Build index
+        print(f"📦 Building {backend_name} index with {len(texts)} documents...")
+        start_time = time.time()
+
+        builder = LeannBuilder(
+            backend_name=backend_name,
+            embedding_model="facebook/contriever",
+            embedding_mode="sentence-transformers",
+            **backend_kwargs,
+        )
+
+        for text in texts:
+            builder.add_text(text)
+
+        builder.build_index(index_path)
+        build_time = time.time() - start_time
+
+        # Measure index size
+        index_dir = Path(index_path).parent
+        index_files = list(index_dir.glob(f"{Path(index_path).stem}.*"))
+        total_size = sum(f.stat().st_size for f in index_files if f.is_file())
+        size_mb = total_size / (1024 * 1024)
+
+        print(f"   ✅ Build completed in {build_time:.2f}s, index size: {size_mb:.1f}MB")
+
+        # Search benchmark
+        print("🔍 Running search benchmark...")
+        searcher = LeannSearcher(index_path)
+
+        search_times = []
+        all_results = []
+
+        for query in test_queries:
+            start_time = time.time()
+            results = searcher.search(query, top_k=5)
+            search_time = time.time() - start_time
+            search_times.append(search_time)
+            all_results.append(results)
+
+        avg_search_time = np.mean(search_times) * 1000  # Convert to ms
+        print(f"   ✅ Average search time: {avg_search_time:.1f}ms")
+
+        # Check for valid scores (detect -inf issues)
+        all_scores = [
+            result.score
+            for results in all_results
+            for result in results
+            if result.score is not None
+        ]
+        valid_scores = [
+            score for score in all_scores if score != float("-inf") and score != float("inf")
+        ]
+        score_validity_rate = len(valid_scores) / len(all_scores) if all_scores else 0
+
+        # Clean up (ensure embedding server shutdown and object GC)
+        try:
+            if hasattr(searcher, "cleanup"):
+                searcher.cleanup()
+            del searcher
+            del builder
+            gc.collect()
+        except Exception as e:
+            print(f"⚠️  Warning: Resource cleanup error: {e}")
+
+        return {
+            "build_time": build_time,
+            "avg_search_time_ms": avg_search_time,
+            "index_size_mb": size_mb,
+            "score_validity_rate": score_validity_rate,
+        }
+
+
+def run_comparison(n_docs: int = 500, n_queries: int = 10):
+    """Run performance comparison between DiskANN and HNSW."""
+    print("🚀 Starting DiskANN vs HNSW Performance Comparison")
+    print(f"📊 Dataset: {n_docs} documents, {n_queries} test queries")
+
+    # Create test data
+    texts = create_test_texts(n_docs)
+    test_queries = [
+        "machine learning algorithms",
+        "natural language processing",
+        "computer vision techniques",
+        "data analysis methods",
+        "neural network architectures",
+        "database query optimization",
+        "software development practices",
+        "security vulnerabilities",
+        "cloud infrastructure",
+        "distributed computing",
+    ][:n_queries]
+
+    # HNSW benchmark
+    hnsw_results = benchmark_backend(
+        backend_name="hnsw",
+        texts=texts,
+        test_queries=test_queries,
+        backend_kwargs={
+            "is_recompute": True,  # Enable recompute for fair comparison
+            "M": 16,
+            "efConstruction": 200,
+        },
+    )
+
+    # DiskANN benchmark
+    diskann_results = benchmark_backend(
+        backend_name="diskann",
+        texts=texts,
+        test_queries=test_queries,
+        backend_kwargs={
+            "is_recompute": True,  # Enable graph partitioning
+            "num_neighbors": 32,
+            "search_list_size": 50,
+        },
+    )
+
+    # Performance comparison
+    print("\n📈 Performance Comparison Results")
+    print(f"{'=' * 60}")
+    print(f"{'Metric':<25} {'HNSW':<15} {'DiskANN':<15} {'Speedup':<10}")
+    print(f"{'-' * 60}")
+
+    # Build time comparison
+    build_speedup = hnsw_results["build_time"] / diskann_results["build_time"]
+    print(
+        f"{'Build Time (s)':<25} {hnsw_results['build_time']:<15.2f} {diskann_results['build_time']:<15.2f} {build_speedup:<10.2f}x"
+    )
+
+    # Search time comparison
+    search_speedup = hnsw_results["avg_search_time_ms"] / diskann_results["avg_search_time_ms"]
+    print(
+        f"{'Search Time (ms)':<25} {hnsw_results['avg_search_time_ms']:<15.1f} {diskann_results['avg_search_time_ms']:<15.1f} {search_speedup:<10.2f}x"
+    )
+
+    # Index size comparison
+    size_ratio = diskann_results["index_size_mb"] / hnsw_results["index_size_mb"]
+    print(
+        f"{'Index Size (MB)':<25} {hnsw_results['index_size_mb']:<15.1f} {diskann_results['index_size_mb']:<15.1f} {size_ratio:<10.2f}x"
+    )
+
+    # Score validity
+    print(
+        f"{'Score Validity (%)':<25} {hnsw_results['score_validity_rate'] * 100:<15.1f} {diskann_results['score_validity_rate'] * 100:<15.1f}"
+    )
+
+    print(f"{'=' * 60}")
+    print("\n🎯 Summary:")
+    if search_speedup > 1:
+        print(f"   DiskANN is {search_speedup:.2f}x faster than HNSW for search")
+    else:
+        print(f"   HNSW is {1 / search_speedup:.2f}x faster than DiskANN for search")
+
+    if size_ratio > 1:
+        print(f"   DiskANN uses {size_ratio:.2f}x more storage than HNSW")
+    else:
+        print(f"   DiskANN uses {1 / size_ratio:.2f}x less storage than HNSW")
+
+    print(
+        f"   Both backends achieved {min(hnsw_results['score_validity_rate'], diskann_results['score_validity_rate']) * 100:.1f}% score validity"
+    )
+
+
+if __name__ == "__main__":
+    import sys
+
+    try:
+        # Handle help request
+        if len(sys.argv) > 1 and sys.argv[1] in ["-h", "--help", "help"]:
+            print("DiskANN vs HNSW Performance Comparison")
+            print("=" * 50)
+            print(f"Usage: python {sys.argv[0]} [n_docs] [n_queries]")
+            print()
+            print("Arguments:")
+            print("  n_docs      Number of documents to index (default: 500)")
+            print("  n_queries   Number of test queries to run (default: 10)")
+            print()
+            print("Examples:")
+            print("  python benchmarks/diskann_vs_hnsw_speed_comparison.py")
+            print("  python benchmarks/diskann_vs_hnsw_speed_comparison.py 1000")
+            print("  python benchmarks/diskann_vs_hnsw_speed_comparison.py 2000 20")
+            sys.exit(0)
+
+        # Parse command line arguments
+        n_docs = int(sys.argv[1]) if len(sys.argv) > 1 else 500
+        n_queries = int(sys.argv[2]) if len(sys.argv) > 2 else 10
+
+        print("DiskANN vs HNSW Performance Comparison")
+        print("=" * 50)
+        print(f"Dataset: {n_docs} documents, {n_queries} queries")
+        print()
+
+        run_comparison(n_docs=n_docs, n_queries=n_queries)
+
+    except KeyboardInterrupt:
+        print("\n⚠️  Benchmark interrupted by user")
+        sys.exit(130)
+    except Exception as e:
+        print(f"\n❌ Benchmark failed: {e}")
+        sys.exit(1)
+    finally:
+        # Ensure clean exit (forceful to prevent rare hangs from atexit/threads)
+        try:
+            gc.collect()
+            print("\n🧹 Cleanup completed")
+            # Flush stdio to ensure message is visible before hard-exit
+            try:
+                import sys as _sys
+
+                _sys.stdout.flush()
+                _sys.stderr.flush()
+            except Exception:
+                pass
+        except Exception:
+            pass
+        # Use os._exit to bypass atexit handlers that may hang in rare cases
+        import os as _os
+
+        _os._exit(0)

benchmarks/enron_emails/README.md

@@ -0,0 +1,141 @@
+# Enron Emails Benchmark
+
+A comprehensive RAG benchmark for evaluating LEANN search and generation on the Enron email corpus. It mirrors the structure and CLI of the existing FinanceBench and LAION benches, using stage-based evaluation with Recall@3 and generation timing.
+
+- Dataset: Enron email CSV (e.g., Kaggle wcukierski/enron-email-dataset) for passages
+- Queries: corbt/enron_emails_sample_questions (filtered for realistic questions)
+- Metrics: Recall@3 vs FAISS Flat baseline + Generation evaluation with Qwen3-8B
+
+## Layout
+
+benchmarks/enron_emails/
+- setup_enron_emails.py: Prepare passages, build LEANN index, build FAISS baseline
+- evaluate_enron_emails.py: Evaluate retrieval recall (Stages 2-5) + generation with Qwen3-8B
+- data/: Generated passages, queries, embeddings-related files
+- baseline/: FAISS Flat baseline files
+- llm_utils.py: LLM utilities for Qwen3-8B generation (in parent directory)
+
+## Quickstart
+
+1) Prepare the data and index
+
+cd benchmarks/enron_emails
+python setup_enron_emails.py --data-dir data
+
+Notes:
+- If `--emails-csv` is omitted, the script attempts to download from Kaggle dataset `wcukierski/enron-email-dataset` using Kaggle API (requires `KAGGLE_USERNAME` and `KAGGLE_KEY`).
+  Alternatively, pass a local path to `--emails-csv`.
+
+Notes:
+- The script parses emails, chunks header/body into passages, builds a compact LEANN index, and then builds a FAISS Flat baseline from the same passages and embedding model.
+- Optionally, it will also create evaluation queries from HuggingFace dataset `corbt/enron_emails_sample_questions`.
+
+2) Run recall evaluation (Stage 2)
+
+python evaluate_enron_emails.py --index data/enron_index_hnsw.leann --stage 2
+
+3) Complexity sweep (Stage 3)
+
+python evaluate_enron_emails.py --index data/enron_index_hnsw.leann --stage 3 --target-recall 0.90 --max-queries 200
+
+Stage 3 uses binary search over complexity to find the minimal value achieving the target Recall@3 (assumes recall is non-decreasing with complexity). The search expands the upper bound as needed and snaps complexity to multiples of 8.
+
+4) Index comparison (Stage 4)
+
+python evaluate_enron_emails.py --index data/enron_index_hnsw.leann --stage 4 --complexity 88 --max-queries 100 --output results.json
+
+5) Generation evaluation (Stage 5)
+
+python evaluate_enron_emails.py --index data/enron_index_hnsw.leann --stage 5 --complexity 88 --llm-backend hf --model-name Qwen/Qwen3-8B
+
+6) Combined index + generation evaluation (Stages 4+5, recommended)
+
+python evaluate_enron_emails.py --index data/enron_index_hnsw.leann --stage 45 --complexity 88 --llm-backend hf
+
+Notes:
+- Minimal CLI: you can run from repo root with only `--index`, defaults match financebench/laion patterns:
+  - `--stage` defaults to `all` (runs 2, 3, 4, 5)
+  - `--baseline-dir` defaults to `baseline`
+  - `--queries` defaults to `data/evaluation_queries.jsonl` (or falls back to the index directory)
+  - `--llm-backend` defaults to `hf` (HuggingFace), can use `vllm`
+  - `--model-name` defaults to `Qwen/Qwen3-8B`
+- Fail-fast behavior: no silent fallbacks. If compact index cannot run with recompute, it errors out.
+- Stage 5 requires Stage 4 retrieval results. Use `--stage 45` to run both efficiently.
+
+Optional flags:
+- --queries data/evaluation_queries.jsonl (custom queries file)
+- --baseline-dir baseline (where FAISS baseline lives)
+- --complexity 88 (LEANN complexity parameter, optimal for 90% recall)
+- --llm-backend hf|vllm (LLM backend for generation)
+- --model-name Qwen/Qwen3-8B (LLM model for generation)
+- --max-queries 1000 (limit number of queries for evaluation)
+
+## Files Produced
+- data/enron_passages_preview.jsonl: Small preview of passages used (for inspection)
+- data/enron_index_hnsw.leann.*: LEANN index files
+- baseline/faiss_flat.index + baseline/metadata.pkl: FAISS baseline with passage IDs
+- data/evaluation_queries.jsonl: Query file (id + query; includes GT IDs for reference)
+
+## Notes
+- Evaluates both retrieval Recall@3 and generation timing with Qwen3-8B thinking model.
+- The emails CSV must contain a column named "message" (raw RFC822 email) and a column named "file" for source identifier. Message-ID headers are parsed as canonical message IDs when present.
+- Qwen3-8B requires special handling for thinking models with chat templates and <think></think> tag processing.
+
+## Stages Summary
+
+- Stage 2 (Recall@3):
+  - Compares LEANN vs FAISS Flat baseline on Recall@3.
+  - Compact index runs with `recompute_embeddings=True`.
+
+- Stage 3 (Binary Search for Complexity):
+  - Builds a non-compact index (`<index>_noncompact.leann`) and runs binary search with `recompute_embeddings=False` to find the minimal complexity achieving target Recall@3 (default 90%).
+
+- Stage 4 (Index Comparison):
+  - Reports .index-only sizes for compact vs non-compact.
+  - Measures timings on queries by default: non-compact (no recompute) vs compact (with recompute).
+  - Stores retrieval results for Stage 5 generation evaluation.
+  - Fails fast if compact recompute cannot run.
+  - If `--complexity` is not provided, the script tries to use the best complexity from Stage 3:
+    - First from the current run (when running `--stage all`), otherwise
+    - From `enron_stage3_results.json` saved next to the index during the last Stage 3 run.
+    - If neither exists, Stage 4 will error and ask you to run Stage 3 or pass `--complexity`.
+
+- Stage 5 (Generation Evaluation):
+  - Uses Qwen3-8B thinking model for RAG generation on retrieved documents from Stage 4.
+  - Supports HuggingFace (`hf`) and vLLM (`vllm`) backends.
+  - Measures generation timing separately from search timing.
+  - Requires Stage 4 results (no additional searching performed).
+
+## Example Results
+
+These are sample results obtained on Enron data using all-mpnet-base-v2 and Qwen3-8B.
+
+- Stage 3 (Binary Search):
+  - Minimal complexity achieving 90% Recall@3: 88
+  - Sampled points:
+    - C=8 → 59.9% Recall@3
+    - C=72 → 89.4% Recall@3
+    - C=88 → 90.2% Recall@3
+    - C=96 → 90.7% Recall@3
+    - C=112 → 91.1% Recall@3
+    - C=136 → 91.3% Recall@3
+    - C=256 → 92.0% Recall@3
+
+- Stage 4 (Index Sizes, .index only):
+  - Compact: ~2.2 MB
+  - Non-compact: ~82.0 MB
+  - Storage saving by compact: ~97.3%
+
+- Stage 4 (Search Timing, 988 queries, complexity=88):
+  - Non-compact (no recompute): ~0.0075 s avg per query
+  - Compact (with recompute): ~1.981 s avg per query
+  - Speed ratio (non-compact/compact): ~0.0038x
+
+- Stage 5 (RAG Generation, 988 queries, Qwen3-8B):
+  - Average generation time: ~22.302 s per query
+  - Total queries processed: 988
+  - LLM backend: HuggingFace transformers
+  - Model: Qwen/Qwen3-8B (thinking model with <think></think> processing)
+
+Full JSON output is saved by the script (see `--output`), e.g.:
+`benchmarks/enron_emails/results_enron_stage45.json`.

benchmarks/enron_emails/data/.gitignore

@@ -0,0 +1 @@
+downloads/

benchmarks/enron_emails/evaluate_enron_emails.py

@@ -0,0 +1,614 @@
+"""
+Enron Emails Benchmark Evaluation - Retrieval Recall@3 (Stages 2/3/4)
+Follows the style of FinanceBench/LAION: Stage 2 recall vs FAISS baseline,
+Stage 3 complexity sweep to target recall, Stage 4 index comparison.
+On errors, fail fast without fallbacks.
+"""
+
+import argparse
+import json
+import logging
+import os
+import pickle
+from pathlib import Path
+
+import numpy as np
+from leann import LeannBuilder, LeannSearcher
+from leann_backend_hnsw import faiss
+
+from ..llm_utils import generate_hf, generate_vllm, load_hf_model, load_vllm_model
+
+# Setup logging to reduce verbose output
+logging.basicConfig(level=logging.WARNING)
+logging.getLogger("leann.api").setLevel(logging.WARNING)
+logging.getLogger("leann_backend_hnsw").setLevel(logging.WARNING)
+
+
+class RecallEvaluator:
+    """Stage 2: Evaluate Recall@3 (LEANN vs FAISS)"""
+
+    def __init__(self, index_path: str, baseline_dir: str):
+        self.index_path = index_path
+        self.baseline_dir = baseline_dir
+        self.searcher = LeannSearcher(index_path)
+
+        baseline_index_path = os.path.join(baseline_dir, "faiss_flat.index")
+        metadata_path = os.path.join(baseline_dir, "metadata.pkl")
+
+        self.faiss_index = faiss.read_index(baseline_index_path)
+        with open(metadata_path, "rb") as f:
+            self.passage_ids = pickle.load(f)
+
+        print(f"📚 Loaded FAISS flat baseline with {self.faiss_index.ntotal} vectors")
+
+        # No fallbacks here; if embedding server is needed but fails, the caller will see the error.
+
+    def evaluate_recall_at_3(
+        self, queries: list[str], complexity: int = 64, recompute_embeddings: bool = True
+    ) -> float:
+        """Evaluate recall@3 using FAISS Flat as ground truth"""
+        from leann.api import compute_embeddings
+
+        recompute_str = "with recompute" if recompute_embeddings else "no recompute"
+        print(f"🔍 Evaluating recall@3 with complexity={complexity} ({recompute_str})...")
+
+        total_recall = 0.0
+        for i, query in enumerate(queries):
+            # Compute query embedding with the same model/mode as the index
+            q_emb = compute_embeddings(
+                [query],
+                self.searcher.embedding_model,
+                mode=self.searcher.embedding_mode,
+                use_server=False,
+            ).astype(np.float32)
+
+            # Search FAISS Flat ground truth
+            n = q_emb.shape[0]
+            k = 3
+            distances = np.zeros((n, k), dtype=np.float32)
+            labels = np.zeros((n, k), dtype=np.int64)
+            self.faiss_index.search(
+                n,
+                faiss.swig_ptr(q_emb),
+                k,
+                faiss.swig_ptr(distances),
+                faiss.swig_ptr(labels),
+            )
+
+            baseline_ids = {self.passage_ids[idx] for idx in labels[0]}
+
+            # Search with LEANN (may require embedding server depending on index configuration)
+            results = self.searcher.search(
+                query,
+                top_k=3,
+                complexity=complexity,
+                recompute_embeddings=recompute_embeddings,
+            )
+            test_ids = {r.id for r in results}
+
+            intersection = test_ids.intersection(baseline_ids)
+            recall = len(intersection) / 3.0
+            total_recall += recall
+
+            if i < 3:
+                print(f"  Q{i + 1}: '{query[:60]}...' -> Recall@3: {recall:.3f}")
+                print(f"    FAISS: {list(baseline_ids)}")
+                print(f"    LEANN: {list(test_ids)}")
+                print(f"    ∩: {list(intersection)}")
+
+        avg = total_recall / max(1, len(queries))
+        print(f"📊 Average Recall@3: {avg:.3f} ({avg * 100:.1f}%)")
+        return avg
+
+    def cleanup(self):
+        if hasattr(self, "searcher"):
+            self.searcher.cleanup()
+
+
+class EnronEvaluator:
+    def __init__(self, index_path: str):
+        self.index_path = index_path
+        self.searcher = LeannSearcher(index_path)
+
+    def load_queries(self, queries_file: str) -> list[str]:
+        queries: list[str] = []
+        with open(queries_file, encoding="utf-8") as f:
+            for line in f:
+                if not line.strip():
+                    continue
+                data = json.loads(line)
+                if "query" in data:
+                    queries.append(data["query"])
+        print(f"📊 Loaded {len(queries)} queries from {queries_file}")
+        return queries
+
+    def cleanup(self):
+        if self.searcher:
+            self.searcher.cleanup()
+
+    def analyze_index_sizes(self) -> dict:
+        """Analyze index sizes (.index only), similar to LAION bench."""
+
+        print("📏 Analyzing index sizes (.index only)...")
+        index_path = Path(self.index_path)
+        index_dir = index_path.parent
+        index_name = index_path.stem
+
+        sizes: dict[str, float] = {}
+        index_file = index_dir / f"{index_name}.index"
+        meta_file = index_dir / f"{index_path.name}.meta.json"
+        passages_file = index_dir / f"{index_path.name}.passages.jsonl"
+        passages_idx_file = index_dir / f"{index_path.name}.passages.idx"
+
+        sizes["index_only_mb"] = (
+            index_file.stat().st_size / (1024 * 1024) if index_file.exists() else 0.0
+        )
+        sizes["metadata_mb"] = (
+            meta_file.stat().st_size / (1024 * 1024) if meta_file.exists() else 0.0
+        )
+        sizes["passages_text_mb"] = (
+            passages_file.stat().st_size / (1024 * 1024) if passages_file.exists() else 0.0
+        )
+        sizes["passages_index_mb"] = (
+            passages_idx_file.stat().st_size / (1024 * 1024) if passages_idx_file.exists() else 0.0
+        )
+
+        print(f"  📁 .index size: {sizes['index_only_mb']:.1f} MB")
+        return sizes
+
+    def create_non_compact_index_for_comparison(self, non_compact_index_path: str) -> dict:
+        """Create a non-compact index for comparison using current passages and embeddings."""
+
+        current_index_path = Path(self.index_path)
+        current_index_dir = current_index_path.parent
+        current_index_name = current_index_path.name
+
+        # Read metadata to get passage source and embedding model
+        meta_path = current_index_dir / f"{current_index_name}.meta.json"
+        with open(meta_path, encoding="utf-8") as f:
+            meta = json.load(f)
+
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not Path(passage_file).is_absolute():
+            passage_file = current_index_dir / Path(passage_file).name
+
+        # Load all passages and ids
+        ids: list[str] = []
+        texts: list[str] = []
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    ids.append(str(data["id"]))
+                    texts.append(data["text"])
+
+        # Compute embeddings using the same method as LEANN
+        from leann.api import compute_embeddings
+
+        embeddings = compute_embeddings(
+            texts,
+            meta["embedding_model"],
+            mode=meta.get("embedding_mode", "sentence-transformers"),
+            use_server=False,
+        ).astype(np.float32)
+
+        # Build non-compact index with same passages and embeddings
+        builder = LeannBuilder(
+            backend_name="hnsw",
+            embedding_model=meta["embedding_model"],
+            embedding_mode=meta.get("embedding_mode", "sentence-transformers"),
+            is_recompute=False,
+            is_compact=False,
+            **{
+                k: v
+                for k, v in meta.get("backend_kwargs", {}).items()
+                if k not in ["is_recompute", "is_compact"]
+            },
+        )
+
+        # Persist a pickle for build_index_from_embeddings
+        pkl_path = current_index_dir / f"{Path(non_compact_index_path).stem}_embeddings.pkl"
+        with open(pkl_path, "wb") as pf:
+            pickle.dump((ids, embeddings), pf)
+
+        print(
+            f"🔨 Building non-compact index at {non_compact_index_path} from precomputed embeddings..."
+        )
+        builder.build_index_from_embeddings(non_compact_index_path, str(pkl_path))
+
+        # Analyze the non-compact index size
+        temp_evaluator = EnronEvaluator(non_compact_index_path)
+        non_compact_sizes = temp_evaluator.analyze_index_sizes()
+        non_compact_sizes["index_type"] = "non_compact"
+
+        return non_compact_sizes
+
+    def compare_index_performance(
+        self, non_compact_path: str, compact_path: str, test_queries: list[str], complexity: int
+    ) -> dict:
+        """Compare search speed for non-compact vs compact indexes."""
+        import time
+
+        results: dict = {
+            "non_compact": {"search_times": []},
+            "compact": {"search_times": []},
+            "avg_search_times": {},
+            "speed_ratio": 0.0,
+            "retrieval_results": [],  # Store retrieval results for Stage 5
+        }
+
+        print("⚡ Comparing search performance between indexes...")
+        # Non-compact (no recompute)
+        print("  🔍 Testing non-compact index (no recompute)...")
+        non_compact_searcher = LeannSearcher(non_compact_path)
+        for q in test_queries:
+            t0 = time.time()
+            _ = non_compact_searcher.search(
+                q, top_k=3, complexity=complexity, recompute_embeddings=False
+            )
+            results["non_compact"]["search_times"].append(time.time() - t0)
+
+        # Compact (with recompute). Fail fast if it cannot run.
+        print("  🔍 Testing compact index (with recompute)...")
+        compact_searcher = LeannSearcher(compact_path)
+        for q in test_queries:
+            t0 = time.time()
+            docs = compact_searcher.search(
+                q, top_k=3, complexity=complexity, recompute_embeddings=True
+            )
+            results["compact"]["search_times"].append(time.time() - t0)
+
+            # Store retrieval results for Stage 5
+            results["retrieval_results"].append(
+                {"query": q, "retrieved_docs": [{"id": doc.id, "text": doc.text} for doc in docs]}
+            )
+        compact_searcher.cleanup()
+
+        if results["non_compact"]["search_times"]:
+            results["avg_search_times"]["non_compact"] = sum(
+                results["non_compact"]["search_times"]
+            ) / len(results["non_compact"]["search_times"])
+        if results["compact"]["search_times"]:
+            results["avg_search_times"]["compact"] = sum(results["compact"]["search_times"]) / len(
+                results["compact"]["search_times"]
+            )
+        if results["avg_search_times"].get("compact", 0) > 0:
+            results["speed_ratio"] = (
+                results["avg_search_times"]["non_compact"] / results["avg_search_times"]["compact"]
+            )
+        else:
+            results["speed_ratio"] = 0.0
+
+        non_compact_searcher.cleanup()
+        return results
+
+    def evaluate_complexity(
+        self,
+        recall_eval: "RecallEvaluator",
+        queries: list[str],
+        target: float = 0.90,
+        c_min: int = 8,
+        c_max: int = 256,
+        max_iters: int = 10,
+        recompute: bool = False,
+    ) -> dict:
+        """Binary search minimal complexity achieving target recall (monotonic assumption)."""
+
+        def round_c(x: int) -> int:
+            # snap to multiple of 8 like other benches typically do
+            return max(1, int((x + 7) // 8) * 8)
+
+        metrics: list[dict] = []
+
+        lo = round_c(c_min)
+        hi = round_c(c_max)
+
+        print(
+            f"🧪 Binary search complexity in [{lo}, {hi}] for target Recall@3>={int(target * 100)}%..."
+        )
+
+        # Ensure upper bound can reach target; expand if needed (up to a cap)
+        r_lo = recall_eval.evaluate_recall_at_3(
+            queries, complexity=lo, recompute_embeddings=recompute
+        )
+        metrics.append({"complexity": lo, "recall_at_3": r_lo})
+        r_hi = recall_eval.evaluate_recall_at_3(
+            queries, complexity=hi, recompute_embeddings=recompute
+        )
+        metrics.append({"complexity": hi, "recall_at_3": r_hi})
+
+        cap = 1024
+        while r_hi < target and hi < cap:
+            lo = hi
+            r_lo = r_hi
+            hi = round_c(hi * 2)
+            r_hi = recall_eval.evaluate_recall_at_3(
+                queries, complexity=hi, recompute_embeddings=recompute
+            )
+            metrics.append({"complexity": hi, "recall_at_3": r_hi})
+
+        if r_hi < target:
+            print(f"⚠️ Max complexity {hi} did not reach target recall {target:.2f}.")
+            print("📈 Observations:")
+            for m in metrics:
+                print(f"  C={m['complexity']:>4} -> Recall@3={m['recall_at_3'] * 100:.1f}%")
+            return {"metrics": metrics, "best_complexity": None, "target_recall": target}
+
+        # Binary search within [lo, hi]
+        best = hi
+        iters = 0
+        while lo < hi and iters < max_iters:
+            mid = round_c((lo + hi) // 2)
+            r_mid = recall_eval.evaluate_recall_at_3(
+                queries, complexity=mid, recompute_embeddings=recompute
+            )
+            metrics.append({"complexity": mid, "recall_at_3": r_mid})
+            if r_mid >= target:
+                best = mid
+                hi = mid
+            else:
+                lo = mid + 8  # move past mid, respecting multiple-of-8 step
+            iters += 1
+
+        print("📈 Binary search results (sampled points):")
+        # Print unique complexity entries ordered by complexity
+        for m in sorted(
+            {m["complexity"]: m for m in metrics}.values(), key=lambda x: x["complexity"]
+        ):
+            print(f"  C={m['complexity']:>4} -> Recall@3={m['recall_at_3'] * 100:.1f}%")
+        print(f"✅ Minimal complexity achieving {int(target * 100)}% recall: {best}")
+        return {"metrics": metrics, "best_complexity": best, "target_recall": target}
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Enron Emails Benchmark Evaluation")
+    parser.add_argument("--index", required=True, help="Path to LEANN index")
+    parser.add_argument(
+        "--queries", default="data/evaluation_queries.jsonl", help="Path to evaluation queries"
+    )
+    parser.add_argument(
+        "--stage",
+        choices=["2", "3", "4", "5", "all", "45"],
+        default="all",
+        help="Which stage to run (2=recall, 3=complexity, 4=index comparison, 5=generation)",
+    )
+    parser.add_argument("--complexity", type=int, default=None, help="LEANN search complexity")
+    parser.add_argument("--baseline-dir", default="baseline", help="Baseline output directory")
+    parser.add_argument(
+        "--max-queries", type=int, help="Limit number of queries to evaluate", default=1000
+    )
+    parser.add_argument(
+        "--target-recall", type=float, default=0.90, help="Target Recall@3 for Stage 3"
+    )
+    parser.add_argument("--output", help="Save results to JSON file")
+    parser.add_argument("--llm-backend", choices=["hf", "vllm"], default="hf", help="LLM backend")
+    parser.add_argument("--model-name", default="Qwen/Qwen3-8B", help="Model name")
+
+    args = parser.parse_args()
+
+    # Resolve queries file: if default path not found, fall back to index's directory
+    if not os.path.exists(args.queries):
+        from pathlib import Path
+
+        idx_dir = Path(args.index).parent
+        fallback_q = idx_dir / "evaluation_queries.jsonl"
+        if fallback_q.exists():
+            args.queries = str(fallback_q)
+
+    baseline_index_path = os.path.join(args.baseline_dir, "faiss_flat.index")
+    if not os.path.exists(baseline_index_path):
+        print(f"❌ FAISS baseline not found at {baseline_index_path}")
+        print("💡 Please run setup_enron_emails.py first to build the baseline")
+        raise SystemExit(1)
+
+    results_out: dict = {}
+
+    if args.stage in ("2", "all"):
+        print("🚀 Starting Stage 2: Recall@3 evaluation")
+        evaluator = RecallEvaluator(args.index, args.baseline_dir)
+
+        enron_eval = EnronEvaluator(args.index)
+        queries = enron_eval.load_queries(args.queries)
+        queries = queries[:10]
+        print(f"🧪 Using first {len(queries)} queries")
+
+        complexity = args.complexity or 64
+        r = evaluator.evaluate_recall_at_3(queries, complexity)
+        results_out["stage2"] = {"complexity": complexity, "recall_at_3": r}
+        evaluator.cleanup()
+        enron_eval.cleanup()
+        print("✅ Stage 2 completed!\n")
+
+    if args.stage in ("3", "all"):
+        print("🚀 Starting Stage 3: Binary search for target recall (no recompute)")
+        enron_eval = EnronEvaluator(args.index)
+        queries = enron_eval.load_queries(args.queries)
+        queries = queries[: args.max_queries]
+        print(f"🧪 Using first {len(queries)} queries")
+
+        # Build non-compact index for fast binary search (recompute_embeddings=False)
+        from pathlib import Path
+
+        index_path = Path(args.index)
+        non_compact_index_path = str(index_path.parent / f"{index_path.stem}_noncompact.leann")
+        enron_eval.create_non_compact_index_for_comparison(non_compact_index_path)
+
+        # Use non-compact evaluator for binary search with recompute=False
+        evaluator_nc = RecallEvaluator(non_compact_index_path, args.baseline_dir)
+        sweep = enron_eval.evaluate_complexity(
+            evaluator_nc, queries, target=args.target_recall, recompute=False
+        )
+        results_out["stage3"] = sweep
+        # Persist default stage 3 results near the index for Stage 4 auto-pickup
+        from pathlib import Path
+
+        default_stage3_path = Path(args.index).parent / "enron_stage3_results.json"
+        with open(default_stage3_path, "w", encoding="utf-8") as f:
+            json.dump({"stage3": sweep}, f, indent=2)
+        print(f"📝 Saved Stage 3 summary to {default_stage3_path}")
+        evaluator_nc.cleanup()
+        enron_eval.cleanup()
+        print("✅ Stage 3 completed!\n")
+
+    if args.stage in ("4", "all", "45"):
+        print("🚀 Starting Stage 4: Index size + performance comparison")
+        evaluator = RecallEvaluator(args.index, args.baseline_dir)
+        enron_eval = EnronEvaluator(args.index)
+        queries = enron_eval.load_queries(args.queries)
+        test_q = queries[: min(args.max_queries, len(queries))]
+
+        current_sizes = enron_eval.analyze_index_sizes()
+        # Build non-compact index for comparison (no fallback)
+        from pathlib import Path
+
+        index_path = Path(args.index)
+        non_compact_path = str(index_path.parent / f"{index_path.stem}_noncompact.leann")
+        non_compact_sizes = enron_eval.create_non_compact_index_for_comparison(non_compact_path)
+        nc_eval = EnronEvaluator(non_compact_path)
+
+        if (
+            current_sizes.get("index_only_mb", 0) > 0
+            and non_compact_sizes.get("index_only_mb", 0) > 0
+        ):
+            storage_saving_percent = max(
+                0.0,
+                100.0 * (1.0 - current_sizes["index_only_mb"] / non_compact_sizes["index_only_mb"]),
+            )
+        else:
+            storage_saving_percent = 0.0
+
+        if args.complexity is None:
+            # Prefer in-session Stage 3 result
+            if "stage3" in results_out and results_out["stage3"].get("best_complexity") is not None:
+                complexity = results_out["stage3"]["best_complexity"]
+                print(f"📥 Using best complexity from Stage 3 in-session: {complexity}")
+            else:
+                # Try to load last saved Stage 3 result near index
+                default_stage3_path = Path(args.index).parent / "enron_stage3_results.json"
+                if default_stage3_path.exists():
+                    with open(default_stage3_path, encoding="utf-8") as f:
+                        prev = json.load(f)
+                    complexity = prev.get("stage3", {}).get("best_complexity")
+                    if complexity is None:
+                        raise SystemExit(
+                            "❌ Stage 4: No --complexity and no best_complexity found in saved Stage 3 results"
+                        )
+                    print(f"📥 Using best complexity from saved Stage 3: {complexity}")
+                else:
+                    raise SystemExit(
+                        "❌ Stage 4 requires --complexity if Stage 3 hasn't been run. Run stage 3 first or pass --complexity."
+                    )
+        else:
+            complexity = args.complexity
+
+        comp = enron_eval.compare_index_performance(
+            non_compact_path, args.index, test_q, complexity=complexity
+        )
+        results_out["stage4"] = {
+            "current_index": current_sizes,
+            "non_compact_index": non_compact_sizes,
+            "storage_saving_percent": storage_saving_percent,
+            "performance_comparison": comp,
+        }
+        nc_eval.cleanup()
+        evaluator.cleanup()
+        enron_eval.cleanup()
+        print("✅ Stage 4 completed!\n")
+
+    if args.stage in ("5", "all"):
+        print("🚀 Starting Stage 5: Generation evaluation with Qwen3-8B")
+
+        # Check if Stage 4 results exist
+        if "stage4" not in results_out or "performance_comparison" not in results_out["stage4"]:
+            print("❌ Stage 5 requires Stage 4 retrieval results")
+            print("💡 Run Stage 4 first or use --stage all")
+            raise SystemExit(1)
+
+        retrieval_results = results_out["stage4"]["performance_comparison"]["retrieval_results"]
+        if not retrieval_results:
+            print("❌ No retrieval results found from Stage 4")
+            raise SystemExit(1)
+
+        print(f"📁 Using {len(retrieval_results)} retrieval results from Stage 4")
+
+        # Load LLM
+        try:
+            if args.llm_backend == "hf":
+                tokenizer, model = load_hf_model(args.model_name)
+
+                def llm_func(prompt):
+                    return generate_hf(tokenizer, model, prompt)
+            else:  # vllm
+                llm, sampling_params = load_vllm_model(args.model_name)
+
+                def llm_func(prompt):
+                    return generate_vllm(llm, sampling_params, prompt)
+
+            # Run generation using stored retrieval results
+            import time
+
+            from llm_utils import create_prompt
+
+            generation_times = []
+            responses = []
+
+            print("🤖 Running generation on pre-retrieved results...")
+            for i, item in enumerate(retrieval_results):
+                query = item["query"]
+                retrieved_docs = item["retrieved_docs"]
+
+                # Prepare context from retrieved docs
+                context = "\n\n".join([doc["text"] for doc in retrieved_docs])
+                prompt = create_prompt(context, query, "emails")
+
+                # Time generation only
+                gen_start = time.time()
+                response = llm_func(prompt)
+                gen_time = time.time() - gen_start
+
+                generation_times.append(gen_time)
+                responses.append(response)
+
+                if i < 3:
+                    print(f"  Q{i + 1}: Gen={gen_time:.3f}s")
+
+            avg_gen_time = sum(generation_times) / len(generation_times)
+
+            print("\n📊 Generation Results:")
+            print(f"  Total Queries: {len(retrieval_results)}")
+            print(f"  Avg Generation Time: {avg_gen_time:.3f}s")
+            print("  (Search time from Stage 4)")
+
+            results_out["stage5"] = {
+                "total_queries": len(retrieval_results),
+                "avg_generation_time": avg_gen_time,
+                "generation_times": generation_times,
+                "responses": responses,
+            }
+
+            # Show sample results
+            print("\n📝 Sample Results:")
+            for i in range(min(3, len(retrieval_results))):
+                query = retrieval_results[i]["query"]
+                response = responses[i]
+                print(f"  Q{i + 1}: {query[:60]}...")
+                print(f"  A{i + 1}: {response[:100]}...")
+                print()
+
+        except Exception as e:
+            print(f"❌ Generation evaluation failed: {e}")
+            print("💡 Make sure transformers/vllm is installed and model is available")
+
+        print("✅ Stage 5 completed!\n")
+
+    if args.output and results_out:
+        with open(args.output, "w", encoding="utf-8") as f:
+            json.dump(results_out, f, indent=2)
+        print(f"📝 Saved results to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/enron_emails/setup_enron_emails.py

@@ -0,0 +1,359 @@
+"""
+Enron Emails Benchmark Setup Script
+Prepares passages from emails.csv, builds LEANN index, and FAISS Flat baseline
+"""
+
+import argparse
+import csv
+import json
+import os
+import re
+from collections.abc import Iterable
+from email import message_from_string
+from email.policy import default
+from pathlib import Path
+from typing import Optional
+
+from leann import LeannBuilder
+
+
+class EnronSetup:
+    def __init__(self, data_dir: str = "data"):
+        self.data_dir = Path(data_dir)
+        self.data_dir.mkdir(parents=True, exist_ok=True)
+
+        self.passages_preview = self.data_dir / "enron_passages_preview.jsonl"
+        self.index_path = self.data_dir / "enron_index_hnsw.leann"
+        self.queries_file = self.data_dir / "evaluation_queries.jsonl"
+        self.downloads_dir = self.data_dir / "downloads"
+        self.downloads_dir.mkdir(parents=True, exist_ok=True)
+
+    # ----------------------------
+    # Dataset acquisition
+    # ----------------------------
+    def ensure_emails_csv(self, emails_csv: Optional[str]) -> str:
+        """Return a path to emails.csv, downloading from Kaggle if needed."""
+        if emails_csv:
+            p = Path(emails_csv)
+            if not p.exists():
+                raise FileNotFoundError(f"emails.csv not found: {emails_csv}")
+            return str(p)
+
+        print(
+            "📥 Trying to download Enron emails.csv from Kaggle (wcukierski/enron-email-dataset)..."
+        )
+        try:
+            from kaggle.api.kaggle_api_extended import KaggleApi
+
+            api = KaggleApi()
+            api.authenticate()
+            api.dataset_download_files(
+                "wcukierski/enron-email-dataset", path=str(self.downloads_dir), unzip=True
+            )
+            candidate = self.downloads_dir / "emails.csv"
+            if candidate.exists():
+                print(f"✅ Downloaded emails.csv: {candidate}")
+                return str(candidate)
+            else:
+                raise FileNotFoundError(
+                    f"emails.csv was not found in {self.downloads_dir} after Kaggle download"
+                )
+        except Exception as e:
+            print(
+                "❌ Could not download via Kaggle automatically. Provide --emails-csv or configure Kaggle API."
+            )
+            print(
+                "   Set KAGGLE_USERNAME and KAGGLE_KEY env vars, or place emails.csv locally and pass --emails-csv."
+            )
+            raise e
+
+    # ----------------------------
+    # Data preparation
+    # ----------------------------
+    @staticmethod
+    def _extract_message_id(raw_email: str) -> str:
+        msg = message_from_string(raw_email, policy=default)
+        val = msg.get("Message-ID", "")
+        if val.startswith("<") and val.endswith(">"):
+            val = val[1:-1]
+        return val or ""
+
+    @staticmethod
+    def _split_header_body(raw_email: str) -> tuple[str, str]:
+        parts = raw_email.split("\n\n", 1)
+        if len(parts) == 2:
+            return parts[0].strip(), parts[1].strip()
+        # Heuristic fallback
+        first_lines = raw_email.splitlines()
+        if first_lines and ":" in first_lines[0]:
+            return raw_email.strip(), ""
+        return "", raw_email.strip()
+
+    @staticmethod
+    def _split_fixed_words(text: str, chunk_words: int, keep_last: bool) -> list[str]:
+        text = (text or "").strip()
+        if not text:
+            return []
+        if chunk_words <= 0:
+            return [text]
+        words = text.split()
+        if not words:
+            return []
+        limit = len(words)
+        if not keep_last:
+            limit = (len(words) // chunk_words) * chunk_words
+        if limit == 0:
+            return []
+        chunks = [" ".join(words[i : i + chunk_words]) for i in range(0, limit, chunk_words)]
+        return [c for c in (s.strip() for s in chunks) if c]
+
+    def _iter_passages_from_csv(
+        self,
+        emails_csv: Path,
+        chunk_words: int = 256,
+        keep_last_header: bool = True,
+        keep_last_body: bool = True,
+        max_emails: int | None = None,
+    ) -> Iterable[dict]:
+        with open(emails_csv, encoding="utf-8") as f:
+            reader = csv.DictReader(f)
+            count = 0
+            for i, row in enumerate(reader):
+                if max_emails is not None and count >= max_emails:
+                    break
+
+                raw_message = row.get("message", "")
+                email_file_id = row.get("file", "")
+
+                if not raw_message.strip():
+                    continue
+
+                message_id = self._extract_message_id(raw_message)
+                if not message_id:
+                    # Fallback ID based on CSV position and file path
+                    safe_file = re.sub(r"[^A-Za-z0-9_.-]", "_", email_file_id)
+                    message_id = f"enron_{i}_{safe_file}"
+
+                header, body = self._split_header_body(raw_message)
+
+                # Header chunks
+                for chunk in self._split_fixed_words(header, chunk_words, keep_last_header):
+                    yield {
+                        "text": chunk,
+                        "metadata": {
+                            "message_id": message_id,
+                            "is_header": True,
+                            "email_file_id": email_file_id,
+                        },
+                    }
+
+                # Body chunks
+                for chunk in self._split_fixed_words(body, chunk_words, keep_last_body):
+                    yield {
+                        "text": chunk,
+                        "metadata": {
+                            "message_id": message_id,
+                            "is_header": False,
+                            "email_file_id": email_file_id,
+                        },
+                    }
+
+                count += 1
+
+    # ----------------------------
+    # Build LEANN index and FAISS baseline
+    # ----------------------------
+    def build_leann_index(
+        self,
+        emails_csv: Optional[str],
+        backend: str = "hnsw",
+        embedding_model: str = "sentence-transformers/all-mpnet-base-v2",
+        chunk_words: int = 256,
+        max_emails: int | None = None,
+    ) -> str:
+        emails_csv_path = self.ensure_emails_csv(emails_csv)
+        print(f"🏗️ Building LEANN index from {emails_csv_path}...")
+
+        builder = LeannBuilder(
+            backend_name=backend,
+            embedding_model=embedding_model,
+            embedding_mode="sentence-transformers",
+            graph_degree=32,
+            complexity=64,
+            is_recompute=True,
+            is_compact=True,
+            num_threads=4,
+        )
+
+        # Stream passages and add to builder
+        preview_written = 0
+        with open(self.passages_preview, "w", encoding="utf-8") as preview_out:
+            for p in self._iter_passages_from_csv(
+                Path(emails_csv_path), chunk_words=chunk_words, max_emails=max_emails
+            ):
+                builder.add_text(p["text"], metadata=p["metadata"])
+                if preview_written < 200:
+                    preview_out.write(json.dumps({"text": p["text"][:200], **p["metadata"]}) + "\n")
+                    preview_written += 1
+
+        print(f"🔨 Building index at {self.index_path}...")
+        builder.build_index(str(self.index_path))
+        print("✅ LEANN index built!")
+        return str(self.index_path)
+
+    def build_faiss_flat_baseline(self, index_path: str, output_dir: str = "baseline") -> str:
+        print("🔨 Building FAISS Flat baseline from LEANN passages...")
+
+        import pickle
+
+        import numpy as np
+        from leann.api import compute_embeddings
+        from leann_backend_hnsw import faiss
+
+        os.makedirs(output_dir, exist_ok=True)
+        baseline_path = os.path.join(output_dir, "faiss_flat.index")
+        metadata_path = os.path.join(output_dir, "metadata.pkl")
+
+        if os.path.exists(baseline_path) and os.path.exists(metadata_path):
+            print(f"✅ Baseline already exists at {baseline_path}")
+            return baseline_path
+
+        # Read meta for passage source and embedding model
+        meta_path = f"{index_path}.meta.json"
+        with open(meta_path, encoding="utf-8") as f:
+            meta = json.load(f)
+
+        embedding_model = meta["embedding_model"]
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        if not os.path.isabs(passage_file):
+            index_dir = os.path.dirname(index_path)
+            passage_file = os.path.join(index_dir, os.path.basename(passage_file))
+
+        # Load passages from builder output so IDs match LEANN
+        passages: list[str] = []
+        passage_ids: list[str] = []
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if not line.strip():
+                    continue
+                data = json.loads(line)
+                passages.append(data["text"])
+                passage_ids.append(data["id"])  # builder-assigned ID
+
+        print(f"📄 Loaded {len(passages)} passages for baseline")
+        print(f"🤖 Embedding model: {embedding_model}")
+
+        embeddings = compute_embeddings(
+            passages,
+            embedding_model,
+            mode="sentence-transformers",
+            use_server=False,
+        )
+
+        # Build FAISS IndexFlatIP
+        dim = embeddings.shape[1]
+        index = faiss.IndexFlatIP(dim)
+        emb_f32 = embeddings.astype(np.float32)
+        index.add(emb_f32.shape[0], faiss.swig_ptr(emb_f32))
+
+        faiss.write_index(index, baseline_path)
+        with open(metadata_path, "wb") as pf:
+            pickle.dump(passage_ids, pf)
+
+        print(f"✅ FAISS baseline saved: {baseline_path}")
+        print(f"✅ Metadata saved: {metadata_path}")
+        print(f"📊 Total vectors: {index.ntotal}")
+        return baseline_path
+
+    # ----------------------------
+    # Queries (optional): prepare evaluation queries file
+    # ----------------------------
+    def prepare_queries(self, min_realism: float = 0.85) -> Path:
+        print(
+            "📝 Preparing evaluation queries from HuggingFace dataset corbt/enron_emails_sample_questions ..."
+        )
+        try:
+            from datasets import load_dataset
+
+            ds = load_dataset("corbt/enron_emails_sample_questions", split="train")
+        except Exception as e:
+            print(f"⚠️  Failed to load dataset: {e}")
+            return self.queries_file
+
+        kept = 0
+        with open(self.queries_file, "w", encoding="utf-8") as out:
+            for i, item in enumerate(ds):
+                how_realistic = float(item.get("how_realistic", 0.0))
+                if how_realistic < min_realism:
+                    continue
+                qid = str(item.get("id", f"enron_q_{i}"))
+                query = item.get("question", "")
+                if not query:
+                    continue
+                record = {
+                    "id": qid,
+                    "query": query,
+                    # For reference only, not used in recall metric below
+                    "gt_message_ids": item.get("message_ids", []),
+                }
+                out.write(json.dumps(record) + "\n")
+                kept += 1
+        print(f"✅ Wrote {kept} queries to {self.queries_file}")
+        return self.queries_file
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Setup Enron Emails Benchmark")
+    parser.add_argument(
+        "--emails-csv",
+        help="Path to emails.csv (Enron dataset). If omitted, attempt Kaggle download.",
+    )
+    parser.add_argument("--data-dir", default="data", help="Data directory")
+    parser.add_argument("--backend", choices=["hnsw", "diskann"], default="hnsw")
+    parser.add_argument(
+        "--embedding-model",
+        default="sentence-transformers/all-mpnet-base-v2",
+        help="Embedding model for LEANN",
+    )
+    parser.add_argument("--chunk-words", type=int, default=256, help="Fixed word chunk size")
+    parser.add_argument("--max-emails", type=int, help="Limit number of emails to process")
+    parser.add_argument("--skip-queries", action="store_true", help="Skip creating queries file")
+    parser.add_argument("--skip-build", action="store_true", help="Skip building LEANN index")
+
+    args = parser.parse_args()
+
+    setup = EnronSetup(args.data_dir)
+
+    # Build index
+    if not args.skip_build:
+        index_path = setup.build_leann_index(
+            emails_csv=args.emails_csv,
+            backend=args.backend,
+            embedding_model=args.embedding_model,
+            chunk_words=args.chunk_words,
+            max_emails=args.max_emails,
+        )
+
+        # Build FAISS baseline from the same passages & embeddings
+        setup.build_faiss_flat_baseline(index_path)
+    else:
+        print("⏭️  Skipping LEANN index build and baseline")
+
+    # Queries file (optional)
+    if not args.skip_queries:
+        setup.prepare_queries()
+    else:
+        print("⏭️  Skipping query preparation")
+
+    print("\n🎉 Enron Emails setup completed!")
+    print(f"📁 Data directory: {setup.data_dir.absolute()}")
+    print("Next steps:")
+    print(
+        "1) Evaluate recall: python evaluate_enron_emails.py --index data/enron_index_hnsw.leann --stage 2"
+    )
+
+
+if __name__ == "__main__":
+    main()

benchmarks/faiss_only.py

@@ -1,11 +1,11 @@
 #!/usr/bin/env python3
 """Test only Faiss HNSW"""
 
+import os
 import sys
 import time
+
 import psutil
-import gc
-import os
 
 
 def get_memory_usage():
@@ -37,20 +37,20 @@ def main():
         import faiss
     except ImportError:
         print("Faiss is not installed.")
-        print("Please install it with `uv pip install faiss-cpu`")
+        print(
+            "Please install it with `uv pip install faiss-cpu` and you can  then run this script again"
+        )
         sys.exit(1)
 
     from llama_index.core import (
-        SimpleDirectoryReader,
-        VectorStoreIndex,
-        StorageContext,
         Settings,
-        node_parser,
-        Document,
+        SimpleDirectoryReader,
+        StorageContext,
+        VectorStoreIndex,
     )
     from llama_index.core.node_parser import SentenceSplitter
-    from llama_index.vector_stores.faiss import FaissVectorStore
     from llama_index.embeddings.huggingface import HuggingFaceEmbedding
+    from llama_index.vector_stores.faiss import FaissVectorStore
 
     tracker = MemoryTracker("Faiss HNSW")
     tracker.checkpoint("Initial")
@@ -65,7 +65,7 @@ def main():
     tracker.checkpoint("After Faiss index creation")
 
     documents = SimpleDirectoryReader(
-        "examples/data",
+        "data",
         recursive=True,
         encoding="utf-8",
         required_exts=[".pdf", ".txt", ".md"],
@@ -90,8 +90,9 @@ def main():
                 vector_store=vector_store, persist_dir="./storage_faiss"
             )
             from llama_index.core import load_index_from_storage
+
             index = load_index_from_storage(storage_context=storage_context)
-            print(f"Index loaded from ./storage_faiss")
+            print("Index loaded from ./storage_faiss")
             tracker.checkpoint("After loading existing index")
             index_loaded = True
         except Exception as e:
@@ -99,19 +100,18 @@ def main():
             print("Cleaning up corrupted index and building new one...")
             # Clean up corrupted index
             import shutil
+
             if os.path.exists("./storage_faiss"):
                 shutil.rmtree("./storage_faiss")
-    
+
     if not index_loaded:
         print("Building new Faiss HNSW index...")
-        
+
         # Use the correct Faiss building pattern from the example
         vector_store = FaissVectorStore(faiss_index=faiss_index)
         storage_context = StorageContext.from_defaults(vector_store=vector_store)
         index = VectorStoreIndex.from_documents(
-            documents, 
-            storage_context=storage_context,
-            transformations=[node_parser]
+            documents, storage_context=storage_context, transformations=[node_parser]
         )
         tracker.checkpoint("After index building")
 
@@ -124,10 +124,10 @@ def main():
     runtime_start_mem = get_memory_usage()
     print(f"Before load memory: {runtime_start_mem:.1f} MB")
     tracker.checkpoint("Before load memory")
-    
+
     query_engine = index.as_query_engine(similarity_top_k=20)
     queries = [
-        "什么是盘古大模型以及盘古开发过程中遇到了什么阴暗面，任务令一般在什么城市颁发",
+        "什么是盘古大模型以及盘古开发过程中遇到了什么阴暗面,任务令一般在什么城市颁发",
         "What is LEANN and how does it work?",
         "华为诺亚方舟实验室的主要研究内容",
     ]
@@ -141,7 +141,7 @@ def main():
 
     runtime_end_mem = get_memory_usage()
     runtime_overhead = runtime_end_mem - runtime_start_mem
-    
+
     peak_memory = tracker.summary()
     print(f"Peak Memory: {peak_memory:.1f} MB")
     print(f"Runtime Memory Overhead: {runtime_overhead:.1f} MB")

benchmarks/financebench/README.md

@@ -0,0 +1,115 @@
+# FinanceBench Benchmark for LEANN-RAG
+
+FinanceBench is a benchmark for evaluating retrieval-augmented generation (RAG) systems on financial document question-answering tasks.
+
+## Dataset
+
+- **Source**: [PatronusAI/financebench](https://huggingface.co/datasets/PatronusAI/financebench)
+- **Questions**: 150 financial Q&A examples
+- **Documents**: 368 PDF files (10-K, 10-Q, 8-K, earnings reports)
+- **Companies**: Major public companies (3M, Apple, Microsoft, Amazon, etc.)
+- **Paper**: [FinanceBench: A New Benchmark for Financial Question Answering](https://arxiv.org/abs/2311.11944)
+
+## Structure
+
+```
+benchmarks/financebench/
+├── setup_financebench.py        # Downloads PDFs and builds index
+├── evaluate_financebench.py     # Intelligent evaluation script
+├── data/
+│   ├── financebench_merged.jsonl     # Q&A dataset
+│   ├── pdfs/                         # Downloaded financial documents
+│   └── index/                        # LEANN indexes
+│       └── financebench_full_hnsw.leann
+└── README.md
+```
+
+## Usage
+
+### 1. Setup (Download & Build Index)
+
+```bash
+cd benchmarks/financebench
+python setup_financebench.py
+```
+
+This will:
+- Download the 150 Q&A examples
+- Download all 368 PDF documents (parallel processing)
+- Build a LEANN index from 53K+ text chunks
+- Verify setup with test query
+
+### 2. Evaluation
+
+```bash
+# Basic retrieval evaluation
+python evaluate_financebench.py --index data/index/financebench_full_hnsw.leann
+
+
+# RAG generation evaluation with Qwen3-8B
+python evaluate_financebench.py --index data/index/financebench_full_hnsw.leann --stage 4 --complexity 64 --llm-backend hf --model-name Qwen/Qwen3-8B --output results_qwen3.json
+```
+
+## Evaluation Methods
+
+### Retrieval Evaluation
+Uses intelligent matching with three strategies:
+1. **Exact text overlap** - Direct substring matches
+2. **Number matching** - Key financial figures ($1,577, 1.2B, etc.)
+3. **Semantic similarity** - Word overlap with 20% threshold
+
+### QA Evaluation
+LLM-based answer evaluation using GPT-4o:
+- Handles numerical rounding and equivalent representations
+- Considers fractions, percentages, and decimal equivalents
+- Evaluates semantic meaning rather than exact text match
+
+## Benchmark Results
+
+### LEANN-RAG Performance (sentence-transformers/all-mpnet-base-v2)
+
+**Retrieval Metrics:**
+- **Question Coverage**: 100.0% (all questions retrieve relevant docs)
+- **Exact Match Rate**: 0.7% (substring overlap with evidence)
+- **Number Match Rate**: 120.7% (key financial figures matched)*
+- **Semantic Match Rate**: 4.7% (word overlap ≥20%)
+- **Average Search Time**: 0.097s
+
+**QA Metrics:**
+- **Accuracy**: 42.7% (LLM-evaluated answer correctness)
+- **Average QA Time**: 4.71s (end-to-end response time)
+
+**System Performance:**
+- **Index Size**: 53,985 chunks from 368 PDFs
+- **Build Time**: ~5-10 minutes with sentence-transformers/all-mpnet-base-v2
+
+*Note: Number match rate >100% indicates multiple retrieved documents contain the same financial figures, which is expected behavior for financial data appearing across multiple document sections.
+
+### LEANN-RAG Generation Performance (Qwen3-8B)
+
+- **Stage 4 (Index Comparison):**
+  - Compact Index: 5.0 MB
+  - Non-compact Index: 172.2 MB
+  - **Storage Saving**: 97.1%
+- **Search Performance**:
+  - Non-compact (no recompute): 0.009s avg per query
+  - Compact (with recompute): 2.203s avg per query
+  - Speed ratio: 0.004x
+
+**Generation Evaluation (20 queries, complexity=64):**
+- **Average Search Time**: 1.638s per query
+- **Average Generation Time**: 45.957s per query
+- **LLM Backend**: HuggingFace transformers
+- **Model**: Qwen/Qwen3-8B (thinking model with <think></think> processing)
+- **Total Questions Processed**: 20
+
+## Options
+
+```bash
+# Use different backends
+python setup_financebench.py --backend diskann
+python evaluate_financebench.py --index data/index/financebench_full_diskann.leann
+
+# Use different embedding models
+python setup_financebench.py --embedding-model facebook/contriever
+```

benchmarks/financebench/evaluate_financebench.py

@@ -0,0 +1,923 @@
+"""
+FinanceBench Evaluation Script - Modular Recall-based Evaluation
+"""
+
+import argparse
+import json
+import logging
+import os
+import pickle
+import time
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import openai
+from leann import LeannChat, LeannSearcher
+from leann_backend_hnsw import faiss
+
+from ..llm_utils import evaluate_rag, generate_hf, generate_vllm, load_hf_model, load_vllm_model
+
+# Setup logging to reduce verbose output
+logging.basicConfig(level=logging.WARNING)
+logging.getLogger("leann.api").setLevel(logging.WARNING)
+logging.getLogger("leann_backend_hnsw").setLevel(logging.WARNING)
+
+
+class RecallEvaluator:
+    """Stage 2: Evaluate Recall@3 (searcher vs baseline)"""
+
+    def __init__(self, index_path: str, baseline_dir: str):
+        self.index_path = index_path
+        self.baseline_dir = baseline_dir
+        self.searcher = LeannSearcher(index_path)
+
+        # Load FAISS flat baseline
+        baseline_index_path = os.path.join(baseline_dir, "faiss_flat.index")
+        metadata_path = os.path.join(baseline_dir, "metadata.pkl")
+
+        self.faiss_index = faiss.read_index(baseline_index_path)
+        with open(metadata_path, "rb") as f:
+            self.passage_ids = pickle.load(f)
+        print(f"📚 Loaded FAISS flat baseline with {self.faiss_index.ntotal} vectors")
+
+    def evaluate_recall_at_3(
+        self, queries: list[str], complexity: int = 64, recompute_embeddings: bool = True
+    ) -> float:
+        """Evaluate recall@3 for given queries at specified complexity"""
+        recompute_str = "with recompute" if recompute_embeddings else "no recompute"
+        print(f"🔍 Evaluating recall@3 with complexity={complexity} ({recompute_str})...")
+
+        total_recall = 0.0
+        num_queries = len(queries)
+
+        for i, query in enumerate(queries):
+            # Get ground truth: search with FAISS flat
+            from leann.api import compute_embeddings
+
+            query_embedding = compute_embeddings(
+                [query],
+                self.searcher.embedding_model,
+                mode=self.searcher.embedding_mode,
+                use_server=False,
+            ).astype(np.float32)
+
+            # Search FAISS flat for ground truth using LEANN's modified faiss API
+            n = query_embedding.shape[0]  # Number of queries
+            k = 3  # Number of nearest neighbors
+            distances = np.zeros((n, k), dtype=np.float32)
+            labels = np.zeros((n, k), dtype=np.int64)
+
+            self.faiss_index.search(
+                n,
+                faiss.swig_ptr(query_embedding),
+                k,
+                faiss.swig_ptr(distances),
+                faiss.swig_ptr(labels),
+            )
+
+            # Extract the results
+            baseline_ids = {self.passage_ids[idx] for idx in labels[0]}
+
+            # Search with LEANN at specified complexity
+            test_results = self.searcher.search(
+                query,
+                top_k=3,
+                complexity=complexity,
+                recompute_embeddings=recompute_embeddings,
+            )
+            test_ids = {result.id for result in test_results}
+
+            # Calculate recall@3 = |intersection| / |ground_truth|
+            intersection = test_ids.intersection(baseline_ids)
+            recall = len(intersection) / 3.0  # Ground truth size is 3
+            total_recall += recall
+
+            if i < 3:  # Show first few examples
+                print(f"  Query {i + 1}: '{query[:50]}...' -> Recall@3: {recall:.3f}")
+                print(f"    FAISS ground truth: {list(baseline_ids)}")
+                print(f"    LEANN results (C={complexity}, {recompute_str}): {list(test_ids)}")
+                print(f"    Intersection: {list(intersection)}")
+
+        avg_recall = total_recall / num_queries
+        print(f"📊 Average Recall@3: {avg_recall:.3f} ({avg_recall * 100:.1f}%)")
+        return avg_recall
+
+    def cleanup(self):
+        """Cleanup resources"""
+        if hasattr(self, "searcher"):
+            self.searcher.cleanup()
+
+
+class FinanceBenchEvaluator:
+    def __init__(self, index_path: str, openai_api_key: Optional[str] = None):
+        self.index_path = index_path
+        self.openai_client = openai.OpenAI(api_key=openai_api_key) if openai_api_key else None
+
+        self.searcher = LeannSearcher(index_path)
+        self.chat = LeannChat(index_path) if openai_api_key else None
+
+    def load_dataset(self, dataset_path: str = "data/financebench_merged.jsonl"):
+        """Load FinanceBench dataset"""
+        data = []
+        with open(dataset_path, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data.append(json.loads(line))
+
+        print(f"📊 Loaded {len(data)} FinanceBench examples")
+        return data
+
+    def analyze_index_sizes(self) -> dict:
+        """Analyze index sizes with and without embeddings"""
+
+        print("📏 Analyzing index sizes...")
+
+        # Get all index-related files
+        index_path = Path(self.index_path)
+        index_dir = index_path.parent
+        index_name = index_path.stem  # Remove .leann extension
+
+        sizes = {}
+        total_with_embeddings = 0
+
+        # Core index files
+        index_file = index_dir / f"{index_name}.index"
+        meta_file = index_dir / f"{index_path.name}.meta.json"  # Keep .leann for meta file
+        passages_file = index_dir / f"{index_path.name}.passages.jsonl"  # Keep .leann for passages
+        passages_idx_file = index_dir / f"{index_path.name}.passages.idx"  # Keep .leann for idx
+
+        for file_path, name in [
+            (index_file, "index"),
+            (meta_file, "metadata"),
+            (passages_file, "passages_text"),
+            (passages_idx_file, "passages_index"),
+        ]:
+            if file_path.exists():
+                size_mb = file_path.stat().st_size / (1024 * 1024)
+                sizes[name] = size_mb
+                total_with_embeddings += size_mb
+
+            else:
+                sizes[name] = 0
+
+        sizes["total_with_embeddings"] = total_with_embeddings
+        sizes["index_only_mb"] = sizes["index"]  # Just the .index file for fair comparison
+
+        print(f"  📁 Total index size: {total_with_embeddings:.1f} MB")
+        print(f"  📁 Index file only: {sizes['index']:.1f} MB")
+
+        return sizes
+
+    def create_compact_index_for_comparison(self, compact_index_path: str) -> dict:
+        """Create a compact index for comparison purposes"""
+        print("🏗️ Building compact index from existing passages...")
+
+        # Load existing passages from current index
+
+        from leann import LeannBuilder
+
+        current_index_path = Path(self.index_path)
+        current_index_dir = current_index_path.parent
+        current_index_name = current_index_path.name
+
+        # Read metadata to get passage source
+        meta_path = current_index_dir / f"{current_index_name}.meta.json"
+        with open(meta_path) as f:
+            import json
+
+            meta = json.load(f)
+
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not Path(passage_file).is_absolute():
+            passage_file = current_index_dir / Path(passage_file).name
+
+        print(f"📄 Loading passages from {passage_file}...")
+
+        # Build compact index with same passages
+        builder = LeannBuilder(
+            backend_name="hnsw",
+            embedding_model=meta["embedding_model"],
+            embedding_mode=meta.get("embedding_mode", "sentence-transformers"),
+            is_recompute=True,  # Enable recompute (no stored embeddings)
+            is_compact=True,  # Enable compact storage
+            **meta.get("backend_kwargs", {}),
+        )
+
+        # Load all passages
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    builder.add_text(data["text"], metadata=data.get("metadata", {}))
+
+        print(f"🔨 Building compact index at {compact_index_path}...")
+        builder.build_index(compact_index_path)
+
+        # Analyze the compact index size
+        temp_evaluator = FinanceBenchEvaluator(compact_index_path)
+        compact_sizes = temp_evaluator.analyze_index_sizes()
+        compact_sizes["index_type"] = "compact"
+
+        return compact_sizes
+
+    def create_non_compact_index_for_comparison(self, non_compact_index_path: str) -> dict:
+        """Create a non-compact index for comparison purposes"""
+        print("🏗️ Building non-compact index from existing passages...")
+
+        # Load existing passages from current index
+
+        from leann import LeannBuilder
+
+        current_index_path = Path(self.index_path)
+        current_index_dir = current_index_path.parent
+        current_index_name = current_index_path.name
+
+        # Read metadata to get passage source
+        meta_path = current_index_dir / f"{current_index_name}.meta.json"
+        with open(meta_path) as f:
+            import json
+
+            meta = json.load(f)
+
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not Path(passage_file).is_absolute():
+            passage_file = current_index_dir / Path(passage_file).name
+
+        print(f"📄 Loading passages from {passage_file}...")
+
+        # Build non-compact index with same passages
+        builder = LeannBuilder(
+            backend_name="hnsw",
+            embedding_model=meta["embedding_model"],
+            embedding_mode=meta.get("embedding_mode", "sentence-transformers"),
+            is_recompute=False,  # Disable recompute (store embeddings)
+            is_compact=False,  # Disable compact storage
+            **{
+                k: v
+                for k, v in meta.get("backend_kwargs", {}).items()
+                if k not in ["is_recompute", "is_compact"]
+            },
+        )
+
+        # Load all passages
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    builder.add_text(data["text"], metadata=data.get("metadata", {}))
+
+        print(f"🔨 Building non-compact index at {non_compact_index_path}...")
+        builder.build_index(non_compact_index_path)
+
+        # Analyze the non-compact index size
+        temp_evaluator = FinanceBenchEvaluator(non_compact_index_path)
+        non_compact_sizes = temp_evaluator.analyze_index_sizes()
+        non_compact_sizes["index_type"] = "non_compact"
+
+        return non_compact_sizes
+
+    def compare_index_performance(
+        self, non_compact_path: str, compact_path: str, test_data: list, complexity: int
+    ) -> dict:
+        """Compare performance between non-compact and compact indexes"""
+        print("⚡ Comparing search performance between indexes...")
+
+        import time
+
+        from leann import LeannSearcher
+
+        # Test queries
+        test_queries = [item["question"] for item in test_data[:5]]
+
+        results = {
+            "non_compact": {"search_times": []},
+            "compact": {"search_times": []},
+            "avg_search_times": {},
+            "speed_ratio": 0.0,
+        }
+
+        # Test non-compact index (no recompute)
+        print("  🔍 Testing non-compact index (no recompute)...")
+        non_compact_searcher = LeannSearcher(non_compact_path)
+
+        for query in test_queries:
+            start_time = time.time()
+            _ = non_compact_searcher.search(
+                query, top_k=3, complexity=complexity, recompute_embeddings=False
+            )
+            search_time = time.time() - start_time
+            results["non_compact"]["search_times"].append(search_time)
+
+        # Test compact index (with recompute)
+        print("  🔍 Testing compact index (with recompute)...")
+        compact_searcher = LeannSearcher(compact_path)
+
+        for query in test_queries:
+            start_time = time.time()
+            _ = compact_searcher.search(
+                query, top_k=3, complexity=complexity, recompute_embeddings=True
+            )
+            search_time = time.time() - start_time
+            results["compact"]["search_times"].append(search_time)
+
+        # Calculate averages
+        results["avg_search_times"]["non_compact"] = sum(
+            results["non_compact"]["search_times"]
+        ) / len(results["non_compact"]["search_times"])
+        results["avg_search_times"]["compact"] = sum(results["compact"]["search_times"]) / len(
+            results["compact"]["search_times"]
+        )
+
+        # Performance ratio
+        if results["avg_search_times"]["compact"] > 0:
+            results["speed_ratio"] = (
+                results["avg_search_times"]["non_compact"] / results["avg_search_times"]["compact"]
+            )
+        else:
+            results["speed_ratio"] = float("inf")
+
+        print(
+            f"    Non-compact (no recompute): {results['avg_search_times']['non_compact']:.3f}s avg"
+        )
+        print(f"    Compact (with recompute): {results['avg_search_times']['compact']:.3f}s avg")
+        print(f"    Speed ratio: {results['speed_ratio']:.2f}x")
+
+        # Cleanup
+        non_compact_searcher.cleanup()
+        compact_searcher.cleanup()
+
+        return results
+
+    def evaluate_timing_breakdown(
+        self, data: list[dict], max_samples: Optional[int] = None
+    ) -> dict:
+        """Evaluate timing breakdown and accuracy by hacking LeannChat.ask() for separated timing"""
+        if not self.chat or not self.openai_client:
+            print("⚠️  Skipping timing evaluation (no OpenAI API key provided)")
+            return {
+                "total_questions": 0,
+                "avg_search_time": 0.0,
+                "avg_generation_time": 0.0,
+                "avg_total_time": 0.0,
+                "accuracy": 0.0,
+            }
+
+        print("🔍🤖 Evaluating timing breakdown and accuracy (search + generation)...")
+
+        if max_samples:
+            data = data[:max_samples]
+            print(f"📝 Using first {max_samples} samples for timing evaluation")
+
+        search_times = []
+        generation_times = []
+        total_times = []
+        correct_answers = 0
+
+        for i, item in enumerate(data):
+            question = item["question"]
+            ground_truth = item["answer"]
+
+            try:
+                # Hack: Monkey-patch the ask method to capture internal timing
+                original_ask = self.chat.ask
+                captured_search_time = None
+                captured_generation_time = None
+
+                def patched_ask(*args, **kwargs):
+                    nonlocal captured_search_time, captured_generation_time
+
+                    # Time the search part
+                    search_start = time.time()
+                    results = self.chat.searcher.search(args[0], top_k=3, complexity=64)
+                    captured_search_time = time.time() - search_start
+
+                    # Time the generation part
+                    context = "\n\n".join([r.text for r in results])
+                    prompt = (
+                        "Here is some retrieved context that might help answer your question:\n\n"
+                        f"{context}\n\n"
+                        f"Question: {args[0]}\n\n"
+                        "Please provide the best answer you can based on this context and your knowledge."
+                    )
+
+                    generation_start = time.time()
+                    answer = self.chat.llm.ask(prompt)
+                    captured_generation_time = time.time() - generation_start
+
+                    return answer
+
+                # Apply the patch
+                self.chat.ask = patched_ask
+
+                # Time the total QA
+                total_start = time.time()
+                generated_answer = self.chat.ask(question)
+                total_time = time.time() - total_start
+
+                # Restore original method
+                self.chat.ask = original_ask
+
+                # Store the timings
+                search_times.append(captured_search_time)
+                generation_times.append(captured_generation_time)
+                total_times.append(total_time)
+
+                # Check accuracy using LLM as judge
+                is_correct = self._check_answer_accuracy(generated_answer, ground_truth, question)
+                if is_correct:
+                    correct_answers += 1
+
+                status = "✅" if is_correct else "❌"
+                print(
+                    f"Question {i + 1}/{len(data)}: {status} Search={captured_search_time:.3f}s, Gen={captured_generation_time:.3f}s, Total={total_time:.3f}s"
+                )
+                print(f"  GT: {ground_truth}")
+                print(f"  Gen: {generated_answer[:100]}...")
+
+            except Exception as e:
+                print(f"  ❌ Error: {e}")
+                search_times.append(0.0)
+                generation_times.append(0.0)
+                total_times.append(0.0)
+
+        accuracy = correct_answers / len(data) if data else 0.0
+
+        metrics = {
+            "total_questions": len(data),
+            "avg_search_time": sum(search_times) / len(search_times) if search_times else 0.0,
+            "avg_generation_time": sum(generation_times) / len(generation_times)
+            if generation_times
+            else 0.0,
+            "avg_total_time": sum(total_times) / len(total_times) if total_times else 0.0,
+            "accuracy": accuracy,
+            "correct_answers": correct_answers,
+            "search_times": search_times,
+            "generation_times": generation_times,
+            "total_times": total_times,
+        }
+
+        return metrics
+
+    def _check_answer_accuracy(
+        self, generated_answer: str, ground_truth: str, question: str
+    ) -> bool:
+        """Check if generated answer matches ground truth using LLM as judge"""
+        judge_prompt = f"""You are an expert judge evaluating financial question answering.
+
+Question: {question}
+
+Ground Truth Answer: {ground_truth}
+
+Generated Answer: {generated_answer}
+
+Task: Determine if the generated answer is factually correct compared to the ground truth. Focus on:
+1. Numerical accuracy (exact values, units, currency)
+2. Key financial concepts and terminology
+3. Overall factual correctness
+
+For financial data, small formatting differences are OK (e.g., "$1,577" vs "1577 million" vs "$1.577 billion"), but the core numerical value must match.
+
+Respond with exactly one word: "CORRECT" if the generated answer is factually accurate, or "INCORRECT" if it's wrong or significantly different."""
+
+        try:
+            judge_response = self.openai_client.chat.completions.create(
+                model="gpt-4o-mini",
+                messages=[{"role": "user", "content": judge_prompt}],
+                max_tokens=10,
+                temperature=0,
+            )
+            judgment = judge_response.choices[0].message.content.strip().upper()
+            return judgment == "CORRECT"
+        except Exception as e:
+            print(f"  ⚠️  Judge error: {e}, falling back to string matching")
+            # Fallback to simple string matching
+            gen_clean = generated_answer.strip().lower().replace("$", "").replace(",", "")
+            gt_clean = ground_truth.strip().lower().replace("$", "").replace(",", "")
+            return gt_clean in gen_clean
+
+    def _print_results(self, timing_metrics: dict):
+        """Print evaluation results"""
+        print("\n🎯 EVALUATION RESULTS")
+        print("=" * 50)
+
+        # Index comparison analysis
+        if "current_index" in timing_metrics and "non_compact_index" in timing_metrics:
+            print("\n📏 Index Comparison Analysis:")
+            current = timing_metrics["current_index"]
+            non_compact = timing_metrics["non_compact_index"]
+
+            print(f"  Compact index (current): {current.get('total_with_embeddings', 0):.1f} MB")
+            print(
+                f"  Non-compact index (with embeddings): {non_compact.get('total_with_embeddings', 0):.1f} MB"
+            )
+            print(
+                f"  Storage saving by compact: {timing_metrics.get('storage_saving_percent', 0):.1f}%"
+            )
+
+            print("  Component breakdown (non-compact):")
+            print(f"    - Main index: {non_compact.get('index', 0):.1f} MB")
+            print(f"    - Passages text: {non_compact.get('passages_text', 0):.1f} MB")
+            print(f"    - Passages index: {non_compact.get('passages_index', 0):.1f} MB")
+            print(f"    - Metadata: {non_compact.get('metadata', 0):.1f} MB")
+
+        # Performance comparison
+        if "performance_comparison" in timing_metrics:
+            perf = timing_metrics["performance_comparison"]
+            print("\n⚡ Performance Comparison:")
+            print(
+                f"  Non-compact (no recompute): {perf.get('avg_search_times', {}).get('non_compact', 0):.3f}s avg"
+            )
+            print(
+                f"  Compact (with recompute): {perf.get('avg_search_times', {}).get('compact', 0):.3f}s avg"
+            )
+            print(f"  Speed ratio: {perf.get('speed_ratio', 0):.2f}x")
+
+        # Legacy single index analysis (fallback)
+        if "total_with_embeddings" in timing_metrics and "current_index" not in timing_metrics:
+            print("\n📏 Index Size Analysis:")
+            print(f"  Total index size: {timing_metrics.get('total_with_embeddings', 0):.1f} MB")
+
+        print("\n📊 Accuracy:")
+        print(f"  Accuracy: {timing_metrics.get('accuracy', 0) * 100:.1f}%")
+        print(
+            f"  Correct Answers: {timing_metrics.get('correct_answers', 0)}/{timing_metrics.get('total_questions', 0)}"
+        )
+
+        print("\n📊 Timing Breakdown:")
+        print(f"  Total Questions: {timing_metrics.get('total_questions', 0)}")
+        print(f"  Avg Search Time: {timing_metrics.get('avg_search_time', 0):.3f}s")
+        print(f"  Avg Generation Time: {timing_metrics.get('avg_generation_time', 0):.3f}s")
+        print(f"  Avg Total Time: {timing_metrics.get('avg_total_time', 0):.3f}s")
+
+        if timing_metrics.get("avg_total_time", 0) > 0:
+            search_pct = (
+                timing_metrics.get("avg_search_time", 0)
+                / timing_metrics.get("avg_total_time", 1)
+                * 100
+            )
+            gen_pct = (
+                timing_metrics.get("avg_generation_time", 0)
+                / timing_metrics.get("avg_total_time", 1)
+                * 100
+            )
+            print("\n📈 Time Distribution:")
+            print(f"  Search: {search_pct:.1f}%")
+            print(f"  Generation: {gen_pct:.1f}%")
+
+    def cleanup(self):
+        """Cleanup resources"""
+        if self.searcher:
+            self.searcher.cleanup()
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Modular FinanceBench Evaluation")
+    parser.add_argument("--index", required=True, help="Path to LEANN index")
+    parser.add_argument("--dataset", default="data/financebench_merged.jsonl", help="Dataset path")
+    parser.add_argument(
+        "--stage",
+        choices=["2", "3", "4", "all"],
+        default="all",
+        help="Which stage to run (2=recall, 3=complexity, 4=generation)",
+    )
+    parser.add_argument("--complexity", type=int, default=None, help="Complexity for search")
+    parser.add_argument("--baseline-dir", default="baseline", help="Baseline output directory")
+    parser.add_argument("--openai-api-key", help="OpenAI API key for generation evaluation")
+    parser.add_argument("--output", help="Save results to JSON file")
+    parser.add_argument(
+        "--llm-backend", choices=["openai", "hf", "vllm"], default="openai", help="LLM backend"
+    )
+    parser.add_argument("--model-name", default="Qwen3-8B", help="Model name for HF/vLLM")
+
+    args = parser.parse_args()
+
+    try:
+        # Check if baseline exists
+        baseline_index_path = os.path.join(args.baseline_dir, "faiss_flat.index")
+        if not os.path.exists(baseline_index_path):
+            print(f"❌ FAISS baseline not found at {baseline_index_path}")
+            print("💡 Please run setup_financebench.py first to build the baseline")
+            exit(1)
+
+        if args.stage == "2" or args.stage == "all":
+            # Stage 2: Recall@3 evaluation
+            print("🚀 Starting Stage 2: Recall@3 evaluation")
+
+            evaluator = RecallEvaluator(args.index, args.baseline_dir)
+
+            # Load FinanceBench queries for testing
+            print("📖 Loading FinanceBench dataset...")
+            queries = []
+            with open(args.dataset, encoding="utf-8") as f:
+                for line in f:
+                    if line.strip():
+                        data = json.loads(line)
+                        queries.append(data["question"])
+
+            # Test with more queries for robust measurement
+            test_queries = queries[:2000]
+            print(f"🧪 Testing with {len(test_queries)} queries")
+
+            # Test with complexity 64
+            complexity = 64
+            recall = evaluator.evaluate_recall_at_3(test_queries, complexity)
+            print(f"📈 Recall@3 at complexity {complexity}: {recall * 100:.1f}%")
+
+            evaluator.cleanup()
+            print("✅ Stage 2 completed!\n")
+
+        # Shared non-compact index path for Stage 3 and 4
+        non_compact_index_path = args.index.replace(".leann", "_noncompact.leann")
+        complexity = args.complexity
+
+        if args.stage == "3" or args.stage == "all":
+            # Stage 3: Binary search for 90% recall complexity (using non-compact index for speed)
+            print("🚀 Starting Stage 3: Binary search for 90% recall complexity")
+            print(
+                "💡 Creating non-compact index for fast binary search with recompute_embeddings=False"
+            )
+
+            # Create non-compact index for binary search (will be reused in Stage 4)
+            print("🏗️ Creating non-compact index for binary search...")
+            evaluator = FinanceBenchEvaluator(args.index)
+            evaluator.create_non_compact_index_for_comparison(non_compact_index_path)
+
+            # Use non-compact index for binary search
+            binary_search_evaluator = RecallEvaluator(non_compact_index_path, args.baseline_dir)
+
+            # Load queries for testing
+            print("📖 Loading FinanceBench dataset...")
+            queries = []
+            with open(args.dataset, encoding="utf-8") as f:
+                for line in f:
+                    if line.strip():
+                        data = json.loads(line)
+                        queries.append(data["question"])
+
+            # Use more queries for robust measurement
+            test_queries = queries[:200]
+            print(f"🧪 Testing with {len(test_queries)} queries")
+
+            # Binary search for 90% recall complexity (without recompute for speed)
+            target_recall = 0.9
+            min_complexity, max_complexity = 1, 32
+
+            print(f"🔍 Binary search for {target_recall * 100}% recall complexity...")
+            print(f"Search range: {min_complexity} to {max_complexity}")
+
+            best_complexity = None
+            best_recall = 0.0
+
+            while min_complexity <= max_complexity:
+                mid_complexity = (min_complexity + max_complexity) // 2
+
+                print(
+                    f"\n🧪 Testing complexity {mid_complexity} (no recompute, non-compact index)..."
+                )
+                # Use recompute_embeddings=False on non-compact index for fast binary search
+                recall = binary_search_evaluator.evaluate_recall_at_3(
+                    test_queries, mid_complexity, recompute_embeddings=False
+                )
+
+                print(
+                    f"  Complexity {mid_complexity}: Recall@3 = {recall:.3f} ({recall * 100:.1f}%)"
+                )
+
+                if recall >= target_recall:
+                    best_complexity = mid_complexity
+                    best_recall = recall
+                    max_complexity = mid_complexity - 1
+                    print("  ✅ Target reached! Searching for lower complexity...")
+                else:
+                    min_complexity = mid_complexity + 1
+                    print("  ❌ Below target. Searching for higher complexity...")
+
+            if best_complexity is not None:
+                print("\n🎯 Optimal complexity found!")
+                print(f"  Complexity: {best_complexity}")
+                print(f"  Recall@3: {best_recall:.3f} ({best_recall * 100:.1f}%)")
+
+                # Test a few complexities around the optimal one for verification
+                print("\n🔬 Verification test around optimal complexity:")
+                verification_complexities = [
+                    max(1, best_complexity - 2),
+                    max(1, best_complexity - 1),
+                    best_complexity,
+                    best_complexity + 1,
+                    best_complexity + 2,
+                ]
+
+                for complexity in verification_complexities:
+                    if complexity <= 512:  # reasonable upper bound
+                        recall = binary_search_evaluator.evaluate_recall_at_3(
+                            test_queries, complexity, recompute_embeddings=False
+                        )
+                        status = "✅" if recall >= target_recall else "❌"
+                        print(f"  {status} Complexity {complexity:3d}: {recall * 100:5.1f}%")
+
+                # Now test the optimal complexity with compact index and recompute for comparison
+                print(
+                    f"\n🔄 Testing optimal complexity {best_complexity} on compact index WITH recompute..."
+                )
+                compact_evaluator = RecallEvaluator(args.index, args.baseline_dir)
+                recall_with_recompute = compact_evaluator.evaluate_recall_at_3(
+                    test_queries[:10], best_complexity, recompute_embeddings=True
+                )
+                print(
+                    f"  ✅ Complexity {best_complexity} (compact index with recompute): {recall_with_recompute * 100:.1f}%"
+                )
+                complexity = best_complexity
+                print(
+                    f"  📊 Recall difference: {abs(best_recall - recall_with_recompute) * 100:.2f}%"
+                )
+                compact_evaluator.cleanup()
+            else:
+                print(f"\n❌ Could not find complexity achieving {target_recall * 100}% recall")
+                print("All tested complexities were below target.")
+
+            # Cleanup evaluators (keep non-compact index for Stage 4)
+            binary_search_evaluator.cleanup()
+            evaluator.cleanup()
+
+            print("✅ Stage 3 completed! Non-compact index saved for Stage 4.\n")
+
+        if args.stage == "4" or args.stage == "all":
+            # Stage 4: Comprehensive evaluation with dual index comparison
+            print("🚀 Starting Stage 4: Comprehensive evaluation with dual index comparison")
+
+            # Use FinanceBench evaluator for QA evaluation
+            evaluator = FinanceBenchEvaluator(
+                args.index, args.openai_api_key if args.llm_backend == "openai" else None
+            )
+
+            print("📖 Loading FinanceBench dataset...")
+            data = evaluator.load_dataset(args.dataset)
+
+            # Step 1: Analyze current (compact) index
+            print("\n📏 Analyzing current index (compact, pruned)...")
+            compact_size_metrics = evaluator.analyze_index_sizes()
+            compact_size_metrics["index_type"] = "compact"
+
+            # Step 2: Use existing non-compact index or create if needed
+            from pathlib import Path
+
+            if Path(non_compact_index_path).exists():
+                print(
+                    f"\n📁 Using existing non-compact index from Stage 3: {non_compact_index_path}"
+                )
+                temp_evaluator = FinanceBenchEvaluator(non_compact_index_path)
+                non_compact_size_metrics = temp_evaluator.analyze_index_sizes()
+                non_compact_size_metrics["index_type"] = "non_compact"
+            else:
+                print("\n🏗️ Creating non-compact index (with embeddings) for comparison...")
+                non_compact_size_metrics = evaluator.create_non_compact_index_for_comparison(
+                    non_compact_index_path
+                )
+
+            # Step 3: Compare index sizes
+            print("\n📊 Index size comparison:")
+            print(
+                f"  Compact index (current): {compact_size_metrics['total_with_embeddings']:.1f} MB"
+            )
+            print(
+                f"  Non-compact index: {non_compact_size_metrics['total_with_embeddings']:.1f} MB"
+            )
+            print("\n📊 Index-only size comparison (.index file only):")
+            print(f"  Compact index: {compact_size_metrics['index_only_mb']:.1f} MB")
+            print(f"  Non-compact index: {non_compact_size_metrics['index_only_mb']:.1f} MB")
+            # Use index-only size for fair comparison (same as Enron emails)
+            storage_saving = (
+                (non_compact_size_metrics["index_only_mb"] - compact_size_metrics["index_only_mb"])
+                / non_compact_size_metrics["index_only_mb"]
+                * 100
+            )
+            print(f"  Storage saving by compact: {storage_saving:.1f}%")
+
+            # Step 4: Performance comparison between the two indexes
+            if complexity is None:
+                raise ValueError("Complexity is required for performance comparison")
+
+            print("\n⚡ Performance comparison between indexes...")
+            performance_metrics = evaluator.compare_index_performance(
+                non_compact_index_path, args.index, data[:10], complexity=complexity
+            )
+
+            # Step 5: Generation evaluation
+            test_samples = 20
+            print(f"\n🧪 Testing with first {test_samples} samples for generation analysis")
+
+            if args.llm_backend == "openai" and args.openai_api_key:
+                print("🔍🤖 Running OpenAI-based generation evaluation...")
+                evaluation_start = time.time()
+                timing_metrics = evaluator.evaluate_timing_breakdown(data[:test_samples])
+                evaluation_time = time.time() - evaluation_start
+            else:
+                print(
+                    f"🔍🤖 Running {args.llm_backend} generation evaluation with {args.model_name}..."
+                )
+                try:
+                    # Load LLM
+                    if args.llm_backend == "hf":
+                        tokenizer, model = load_hf_model(args.model_name)
+
+                        def llm_func(prompt):
+                            return generate_hf(tokenizer, model, prompt)
+                    else:  # vllm
+                        llm, sampling_params = load_vllm_model(args.model_name)
+
+                        def llm_func(prompt):
+                            return generate_vllm(llm, sampling_params, prompt)
+
+                    # Simple generation evaluation
+                    queries = [item["question"] for item in data[:test_samples]]
+                    gen_results = evaluate_rag(
+                        evaluator.searcher,
+                        llm_func,
+                        queries,
+                        domain="finance",
+                        complexity=complexity,
+                    )
+
+                    timing_metrics = {
+                        "total_questions": len(queries),
+                        "avg_search_time": gen_results["avg_search_time"],
+                        "avg_generation_time": gen_results["avg_generation_time"],
+                        "results": gen_results["results"],
+                    }
+                    evaluation_time = time.time()
+
+                except Exception as e:
+                    print(f"❌ Generation evaluation failed: {e}")
+                    timing_metrics = {
+                        "total_questions": 0,
+                        "avg_search_time": 0,
+                        "avg_generation_time": 0,
+                    }
+                    evaluation_time = 0
+
+            # Combine all metrics
+            combined_metrics = {
+                **timing_metrics,
+                "total_evaluation_time": evaluation_time,
+                "current_index": compact_size_metrics,
+                "non_compact_index": non_compact_size_metrics,
+                "performance_comparison": performance_metrics,
+                "storage_saving_percent": storage_saving,
+            }
+
+            # Print results
+            print("\n📊 Generation Results:")
+            print(f"  Total Questions: {timing_metrics.get('total_questions', 0)}")
+            print(f"  Avg Search Time: {timing_metrics.get('avg_search_time', 0):.3f}s")
+            print(f"  Avg Generation Time: {timing_metrics.get('avg_generation_time', 0):.3f}s")
+
+            # Save results if requested
+            if args.output:
+                print(f"\n💾 Saving results to {args.output}...")
+                with open(args.output, "w") as f:
+                    json.dump(combined_metrics, f, indent=2, default=str)
+                print(f"✅ Results saved to {args.output}")
+
+            evaluator.cleanup()
+            print("✅ Stage 4 completed!\n")
+
+        if args.stage == "all":
+            print("🎉 All evaluation stages completed successfully!")
+            print("\n📋 Summary:")
+            print("  Stage 2: ✅ Recall@3 evaluation completed")
+            print("  Stage 3: ✅ Optimal complexity found")
+            print("  Stage 4: ✅ Generation accuracy & timing evaluation completed")
+            print("\n🔧 Recommended next steps:")
+            print("  - Use optimal complexity for best speed/accuracy balance")
+            print("  - Review accuracy and timing breakdown for performance optimization")
+            print("  - Run full evaluation on complete dataset if needed")
+
+            # Clean up non-compact index after all stages complete
+            print("\n🧹 Cleaning up temporary non-compact index...")
+            from pathlib import Path
+
+            if Path(non_compact_index_path).exists():
+                temp_index_dir = Path(non_compact_index_path).parent
+                temp_index_name = Path(non_compact_index_path).name
+                for temp_file in temp_index_dir.glob(f"{temp_index_name}*"):
+                    temp_file.unlink()
+                print(f"✅ Cleaned up {non_compact_index_path}")
+            else:
+                print("📝 No temporary index to clean up")
+    except KeyboardInterrupt:
+        print("\n⚠️  Evaluation interrupted by user")
+        exit(1)
+    except Exception as e:
+        print(f"\n❌ Stage {args.stage} failed: {e}")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/financebench/setup_financebench.py

@@ -0,0 +1,462 @@
+#!/usr/bin/env python3
+"""
+FinanceBench Complete Setup Script
+Downloads all PDFs and builds full LEANN datastore
+"""
+
+import argparse
+import os
+import re
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+from threading import Lock
+
+import pymupdf
+import requests
+from leann import LeannBuilder, LeannSearcher
+from tqdm import tqdm
+
+
+class FinanceBenchSetup:
+    def __init__(self, data_dir: str = "data"):
+        self.base_dir = Path(__file__).parent  # benchmarks/financebench/
+        self.data_dir = self.base_dir / data_dir
+        self.pdf_dir = self.data_dir / "pdfs"
+        self.dataset_file = self.data_dir / "financebench_merged.jsonl"
+        self.index_dir = self.data_dir / "index"
+        self.download_lock = Lock()
+
+    def download_dataset(self):
+        """Download the main FinanceBench dataset"""
+        print("📊 Downloading FinanceBench dataset...")
+        self.data_dir.mkdir(parents=True, exist_ok=True)
+
+        if self.dataset_file.exists():
+            print(f"✅ Dataset already exists: {self.dataset_file}")
+            return
+
+        url = "https://huggingface.co/datasets/PatronusAI/financebench/raw/main/financebench_merged.jsonl"
+        response = requests.get(url, stream=True)
+        response.raise_for_status()
+
+        with open(self.dataset_file, "wb") as f:
+            for chunk in response.iter_content(chunk_size=8192):
+                f.write(chunk)
+
+        print(f"✅ Dataset downloaded: {self.dataset_file}")
+
+    def get_pdf_list(self):
+        """Get list of all PDF files from GitHub"""
+        print("📋 Fetching PDF list from GitHub...")
+
+        response = requests.get(
+            "https://api.github.com/repos/patronus-ai/financebench/contents/pdfs"
+        )
+        response.raise_for_status()
+        pdf_files = response.json()
+
+        print(f"Found {len(pdf_files)} PDF files")
+        return pdf_files
+
+    def download_single_pdf(self, pdf_info, position):
+        """Download a single PDF file"""
+        pdf_name = pdf_info["name"]
+        pdf_path = self.pdf_dir / pdf_name
+
+        # Skip if already downloaded
+        if pdf_path.exists() and pdf_path.stat().st_size > 0:
+            return f"✅ {pdf_name} (cached)"
+
+        try:
+            # Download PDF
+            response = requests.get(pdf_info["download_url"], timeout=60)
+            response.raise_for_status()
+
+            # Write to file
+            with self.download_lock:
+                with open(pdf_path, "wb") as f:
+                    f.write(response.content)
+
+            return f"✅ {pdf_name} ({len(response.content) // 1024}KB)"
+
+        except Exception as e:
+            return f"❌ {pdf_name}: {e!s}"
+
+    def download_all_pdfs(self, max_workers: int = 5):
+        """Download all PDF files with parallel processing"""
+        self.pdf_dir.mkdir(parents=True, exist_ok=True)
+
+        pdf_files = self.get_pdf_list()
+
+        print(f"📥 Downloading {len(pdf_files)} PDFs with {max_workers} workers...")
+
+        with ThreadPoolExecutor(max_workers=max_workers) as executor:
+            # Submit all download tasks
+            future_to_pdf = {
+                executor.submit(self.download_single_pdf, pdf_info, i): pdf_info["name"]
+                for i, pdf_info in enumerate(pdf_files)
+            }
+
+            # Process completed downloads with progress bar
+            with tqdm(total=len(pdf_files), desc="Downloading PDFs") as pbar:
+                for future in as_completed(future_to_pdf):
+                    result = future.result()
+                    pbar.set_postfix_str(result.split()[-1] if "✅" in result else "Error")
+                    pbar.update(1)
+
+        # Verify downloads
+        downloaded_pdfs = list(self.pdf_dir.glob("*.pdf"))
+        print(f"✅ Successfully downloaded {len(downloaded_pdfs)}/{len(pdf_files)} PDFs")
+
+        # Show any failures
+        missing_pdfs = []
+        for pdf_info in pdf_files:
+            pdf_path = self.pdf_dir / pdf_info["name"]
+            if not pdf_path.exists() or pdf_path.stat().st_size == 0:
+                missing_pdfs.append(pdf_info["name"])
+
+        if missing_pdfs:
+            print(f"⚠️  Failed to download {len(missing_pdfs)} PDFs:")
+            for pdf in missing_pdfs[:5]:  # Show first 5
+                print(f"   - {pdf}")
+            if len(missing_pdfs) > 5:
+                print(f"   ... and {len(missing_pdfs) - 5} more")
+
+    def build_leann_index(
+        self,
+        backend: str = "hnsw",
+        embedding_model: str = "sentence-transformers/all-mpnet-base-v2",
+    ):
+        """Build LEANN index from all PDFs"""
+        print(f"🏗️  Building LEANN index with {backend} backend...")
+
+        # Check if we have PDFs
+        pdf_files = list(self.pdf_dir.glob("*.pdf"))
+        if not pdf_files:
+            raise RuntimeError("No PDF files found! Run download first.")
+
+        print(f"Found {len(pdf_files)} PDF files to process")
+
+        start_time = time.time()
+
+        # Initialize builder with standard compact configuration
+        builder = LeannBuilder(
+            backend_name=backend,
+            embedding_model=embedding_model,
+            embedding_mode="sentence-transformers",
+            graph_degree=32,
+            complexity=64,
+            is_recompute=True,  # Enable recompute (no stored embeddings)
+            is_compact=True,  # Enable compact storage (pruned)
+            num_threads=4,
+        )
+
+        # Process PDFs and extract text
+        total_chunks = 0
+        failed_pdfs = []
+
+        for pdf_path in tqdm(pdf_files, desc="Processing PDFs"):
+            try:
+                chunks = self.extract_pdf_text(pdf_path)
+                for chunk in chunks:
+                    builder.add_text(chunk["text"], metadata=chunk["metadata"])
+                    total_chunks += 1
+
+            except Exception as e:
+                print(f"❌ Failed to process {pdf_path.name}: {e}")
+                failed_pdfs.append(pdf_path.name)
+                continue
+
+        # Build index in index directory
+        self.index_dir.mkdir(parents=True, exist_ok=True)
+        index_path = self.index_dir / f"financebench_full_{backend}.leann"
+        print(f"🔨 Building index: {index_path}")
+        builder.build_index(str(index_path))
+
+        build_time = time.time() - start_time
+
+        print("✅ Index built successfully!")
+        print(f"   📁 Index path: {index_path}")
+        print(f"   📊 Total chunks: {total_chunks:,}")
+        print(f"   📄 Processed PDFs: {len(pdf_files) - len(failed_pdfs)}/{len(pdf_files)}")
+        print(f"   ⏱️  Build time: {build_time:.1f}s")
+
+        if failed_pdfs:
+            print(f"   ⚠️  Failed PDFs: {failed_pdfs}")
+
+        return str(index_path)
+
+    def build_faiss_flat_baseline(self, index_path: str, output_dir: str = "baseline"):
+        """Build FAISS flat baseline using the same embeddings as LEANN index"""
+        print("🔨 Building FAISS Flat baseline...")
+
+        import os
+        import pickle
+
+        import numpy as np
+        from leann.api import compute_embeddings
+        from leann_backend_hnsw import faiss
+
+        os.makedirs(output_dir, exist_ok=True)
+        baseline_path = os.path.join(output_dir, "faiss_flat.index")
+        metadata_path = os.path.join(output_dir, "metadata.pkl")
+
+        if os.path.exists(baseline_path) and os.path.exists(metadata_path):
+            print(f"✅ Baseline already exists at {baseline_path}")
+            return baseline_path
+
+        # Read metadata from the built index
+        meta_path = f"{index_path}.meta.json"
+        with open(meta_path) as f:
+            import json
+
+            meta = json.loads(f.read())
+
+        embedding_model = meta["embedding_model"]
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not os.path.isabs(passage_file):
+            index_dir = os.path.dirname(index_path)
+            passage_file = os.path.join(index_dir, os.path.basename(passage_file))
+
+        print(f"📊 Loading passages from {passage_file}...")
+        print(f"🤖 Using embedding model: {embedding_model}")
+
+        # Load all passages for baseline
+        passages = []
+        passage_ids = []
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    passages.append(data["text"])
+                    passage_ids.append(data["id"])
+
+        print(f"📄 Loaded {len(passages)} passages")
+
+        # Compute embeddings using the same method as LEANN
+        print("🧮 Computing embeddings...")
+        embeddings = compute_embeddings(
+            passages,
+            embedding_model,
+            mode="sentence-transformers",
+            use_server=False,
+        )
+
+        print(f"📐 Embedding shape: {embeddings.shape}")
+
+        # Build FAISS flat index
+        print("🏗️  Building FAISS IndexFlatIP...")
+        dimension = embeddings.shape[1]
+        index = faiss.IndexFlatIP(dimension)
+
+        # Add embeddings to flat index
+        embeddings_f32 = embeddings.astype(np.float32)
+        index.add(embeddings_f32.shape[0], faiss.swig_ptr(embeddings_f32))
+
+        # Save index and metadata
+        faiss.write_index(index, baseline_path)
+        with open(metadata_path, "wb") as f:
+            pickle.dump(passage_ids, f)
+
+        print(f"✅ FAISS baseline saved to {baseline_path}")
+        print(f"✅ Metadata saved to {metadata_path}")
+        print(f"📊 Total vectors: {index.ntotal}")
+
+        return baseline_path
+
+    def extract_pdf_text(self, pdf_path: Path) -> list[dict]:
+        """Extract and chunk text from a PDF file"""
+        chunks = []
+        doc = pymupdf.open(pdf_path)
+
+        for page_num in range(len(doc)):
+            page = doc[page_num]
+            text = page.get_text()  # type: ignore
+
+            if not text.strip():
+                continue
+
+            # Create metadata
+            metadata = {
+                "source_file": pdf_path.name,
+                "page_number": page_num + 1,
+                "document_type": "10K" if "10K" in pdf_path.name else "10Q",
+                "company": pdf_path.name.split("_")[0],
+                "doc_period": self.extract_year_from_filename(pdf_path.name),
+            }
+
+            # Use recursive character splitting like LangChain
+            if len(text.split()) > 500:
+                # Split by double newlines (paragraphs)
+                paragraphs = [p.strip() for p in text.split("\n\n") if p.strip()]
+
+                current_chunk = ""
+                for para in paragraphs:
+                    # If adding this paragraph would make chunk too long, save current chunk
+                    if current_chunk and len((current_chunk + " " + para).split()) > 300:
+                        if current_chunk.strip():
+                            chunks.append(
+                                {
+                                    "text": current_chunk.strip(),
+                                    "metadata": {
+                                        **metadata,
+                                        "chunk_id": f"page_{page_num + 1}_chunk_{len(chunks)}",
+                                    },
+                                }
+                            )
+                        current_chunk = para
+                    else:
+                        current_chunk = (current_chunk + " " + para).strip()
+
+                # Add the last chunk
+                if current_chunk.strip():
+                    chunks.append(
+                        {
+                            "text": current_chunk.strip(),
+                            "metadata": {
+                                **metadata,
+                                "chunk_id": f"page_{page_num + 1}_chunk_{len(chunks)}",
+                            },
+                        }
+                    )
+            else:
+                # Page is short enough, use as single chunk
+                chunks.append(
+                    {
+                        "text": text.strip(),
+                        "metadata": {**metadata, "chunk_id": f"page_{page_num + 1}"},
+                    }
+                )
+
+        doc.close()
+        return chunks
+
+    def extract_year_from_filename(self, filename: str) -> str:
+        """Extract year from PDF filename"""
+        # Try to find 4-digit year in filename
+
+        match = re.search(r"(\d{4})", filename)
+        return match.group(1) if match else "unknown"
+
+    def verify_setup(self, index_path: str):
+        """Verify the setup by testing a simple query"""
+        print("🧪 Verifying setup with test query...")
+
+        try:
+            searcher = LeannSearcher(index_path)
+
+            # Test query
+            test_query = "What is the capital expenditure for 3M in 2018?"
+            results = searcher.search(test_query, top_k=3)
+
+            print(f"✅ Test query successful! Found {len(results)} results:")
+            for i, result in enumerate(results, 1):
+                company = result.metadata.get("company", "Unknown")
+                year = result.metadata.get("doc_period", "Unknown")
+                page = result.metadata.get("page_number", "Unknown")
+                print(f"   {i}. {company} {year} (page {page}) - Score: {result.score:.3f}")
+                print(f"      {result.text[:100]}...")
+
+            searcher.cleanup()
+            print("✅ Setup verification completed successfully!")
+
+        except Exception as e:
+            print(f"❌ Setup verification failed: {e}")
+            raise
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Setup FinanceBench with full PDF datastore")
+    parser.add_argument("--data-dir", default="data", help="Data directory")
+    parser.add_argument(
+        "--backend", choices=["hnsw", "diskann"], default="hnsw", help="LEANN backend"
+    )
+    parser.add_argument(
+        "--embedding-model",
+        default="sentence-transformers/all-mpnet-base-v2",
+        help="Embedding model",
+    )
+    parser.add_argument("--max-workers", type=int, default=5, help="Parallel download workers")
+    parser.add_argument("--skip-download", action="store_true", help="Skip PDF download")
+    parser.add_argument("--skip-build", action="store_true", help="Skip index building")
+    parser.add_argument(
+        "--build-baseline-only",
+        action="store_true",
+        help="Only build FAISS baseline from existing index",
+    )
+
+    args = parser.parse_args()
+
+    print("🏦 FinanceBench Complete Setup")
+    print("=" * 50)
+
+    setup = FinanceBenchSetup(args.data_dir)
+
+    try:
+        if args.build_baseline_only:
+            # Only build baseline from existing index
+            index_path = setup.index_dir / f"financebench_full_{args.backend}"
+            index_file = f"{index_path}.index"
+            meta_file = f"{index_path}.leann.meta.json"
+
+            if not os.path.exists(index_file) or not os.path.exists(meta_file):
+                print("❌ Index files not found:")
+                print(f"   Index: {index_file}")
+                print(f"   Meta: {meta_file}")
+                print("💡 Run without --build-baseline-only to build the index first")
+                exit(1)
+
+            print(f"🔨 Building baseline from existing index: {index_path}")
+            baseline_path = setup.build_faiss_flat_baseline(str(index_path))
+            print(f"✅ Baseline built at {baseline_path}")
+            return
+
+        # Step 1: Download dataset
+        setup.download_dataset()
+
+        # Step 2: Download PDFs
+        if not args.skip_download:
+            setup.download_all_pdfs(max_workers=args.max_workers)
+        else:
+            print("⏭️  Skipping PDF download")
+
+        # Step 3: Build LEANN index
+        if not args.skip_build:
+            index_path = setup.build_leann_index(
+                backend=args.backend, embedding_model=args.embedding_model
+            )
+
+            # Step 4: Build FAISS flat baseline
+            print("\n🔨 Building FAISS flat baseline...")
+            baseline_path = setup.build_faiss_flat_baseline(index_path)
+            print(f"✅ Baseline built at {baseline_path}")
+
+            # Step 5: Verify setup
+            setup.verify_setup(index_path)
+        else:
+            print("⏭️  Skipping index building")
+
+        print("\n🎉 FinanceBench setup completed!")
+        print(f"📁 Data directory: {setup.data_dir.absolute()}")
+        print("\nNext steps:")
+        print(
+            "1. Run evaluation: python evaluate_financebench.py --index data/index/financebench_full_hnsw.leann"
+        )
+        print(
+            "2. Or test manually: python -c \"from leann import LeannSearcher; s = LeannSearcher('data/index/financebench_full_hnsw.leann'); print(s.search('3M capital expenditure 2018'))\""
+        )
+
+    except KeyboardInterrupt:
+        print("\n⚠️  Setup interrupted by user")
+        exit(1)
+    except Exception as e:
+        print(f"\n❌ Setup failed: {e}")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/financebench/verify_recall.py

@@ -0,0 +1,214 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.9"
+# dependencies = [
+#     "faiss-cpu",
+#     "numpy",
+#     "sentence-transformers",
+#     "torch",
+#     "tqdm",
+# ]
+# ///
+
+"""
+Independent recall verification script using standard FAISS.
+Creates two indexes (HNSW and Flat) and compares recall@3 at different complexities.
+"""
+
+import json
+import time
+from pathlib import Path
+
+import faiss
+import numpy as np
+from sentence_transformers import SentenceTransformer
+from tqdm import tqdm
+
+
+def compute_embeddings_direct(chunks: list[str], model_name: str) -> np.ndarray:
+    """
+    Direct embedding computation using sentence-transformers.
+    Copied logic to avoid dependency issues.
+    """
+    print(f"Loading model: {model_name}")
+    model = SentenceTransformer(model_name)
+
+    print(f"Computing embeddings for {len(chunks)} chunks...")
+    embeddings = model.encode(
+        chunks,
+        show_progress_bar=True,
+        batch_size=32,
+        convert_to_numpy=True,
+        normalize_embeddings=False,
+    )
+
+    return embeddings.astype(np.float32)
+
+
+def load_financebench_queries(dataset_path: str, max_queries: int = 200) -> list[str]:
+    """Load FinanceBench queries from dataset"""
+    queries = []
+    with open(dataset_path, encoding="utf-8") as f:
+        for line in f:
+            if line.strip():
+                data = json.loads(line)
+                queries.append(data["question"])
+                if len(queries) >= max_queries:
+                    break
+    return queries
+
+
+def load_passages_from_leann_index(index_path: str) -> tuple[list[str], list[str]]:
+    """Load passages from LEANN index structure"""
+    meta_path = f"{index_path}.meta.json"
+    with open(meta_path) as f:
+        meta = json.load(f)
+
+    passage_source = meta["passage_sources"][0]
+    passage_file = passage_source["path"]
+
+    # Convert relative path to absolute
+    if not Path(passage_file).is_absolute():
+        index_dir = Path(index_path).parent
+        passage_file = index_dir / Path(passage_file).name
+
+    print(f"Loading passages from {passage_file}")
+
+    passages = []
+    passage_ids = []
+    with open(passage_file, encoding="utf-8") as f:
+        for line in tqdm(f, desc="Loading passages"):
+            if line.strip():
+                data = json.loads(line)
+                passages.append(data["text"])
+                passage_ids.append(data["id"])
+
+    print(f"Loaded {len(passages)} passages")
+    return passages, passage_ids
+
+
+def build_faiss_indexes(embeddings: np.ndarray) -> tuple[faiss.Index, faiss.Index]:
+    """Build FAISS indexes: Flat (ground truth) and HNSW"""
+    dimension = embeddings.shape[1]
+
+    # Build Flat index (ground truth)
+    print("Building FAISS IndexFlatIP (ground truth)...")
+    flat_index = faiss.IndexFlatIP(dimension)
+    flat_index.add(embeddings)
+
+    # Build HNSW index
+    print("Building FAISS IndexHNSWFlat...")
+    M = 32  # Same as LEANN default
+    hnsw_index = faiss.IndexHNSWFlat(dimension, M, faiss.METRIC_INNER_PRODUCT)
+    hnsw_index.hnsw.efConstruction = 200  # Same as LEANN default
+    hnsw_index.add(embeddings)
+
+    print(f"Built indexes with {flat_index.ntotal} vectors, dimension {dimension}")
+    return flat_index, hnsw_index
+
+
+def evaluate_recall_at_k(
+    query_embeddings: np.ndarray,
+    flat_index: faiss.Index,
+    hnsw_index: faiss.Index,
+    passage_ids: list[str],
+    k: int = 3,
+    ef_search: int = 64,
+) -> float:
+    """Evaluate recall@k comparing HNSW vs Flat"""
+
+    # Set search parameters for HNSW
+    hnsw_index.hnsw.efSearch = ef_search
+
+    total_recall = 0.0
+    num_queries = query_embeddings.shape[0]
+
+    for i in range(num_queries):
+        query = query_embeddings[i : i + 1]  # Keep 2D shape
+
+        # Get ground truth from Flat index (standard FAISS API)
+        flat_distances, flat_indices = flat_index.search(query, k)
+        ground_truth_ids = {passage_ids[idx] for idx in flat_indices[0]}
+
+        # Get results from HNSW index (standard FAISS API)
+        hnsw_distances, hnsw_indices = hnsw_index.search(query, k)
+        hnsw_ids = {passage_ids[idx] for idx in hnsw_indices[0]}
+
+        # Calculate recall
+        intersection = ground_truth_ids.intersection(hnsw_ids)
+        recall = len(intersection) / k
+        total_recall += recall
+
+        if i < 3:  # Show first few examples
+            print(f"  Query {i + 1}: Recall@{k} = {recall:.3f}")
+            print(f"    Flat: {list(ground_truth_ids)}")
+            print(f"    HNSW: {list(hnsw_ids)}")
+            print(f"    Intersection: {list(intersection)}")
+
+    avg_recall = total_recall / num_queries
+    return avg_recall
+
+
+def main():
+    # Configuration
+    dataset_path = "data/financebench_merged.jsonl"
+    index_path = "data/index/financebench_full_hnsw.leann"
+    embedding_model = "sentence-transformers/all-mpnet-base-v2"
+
+    print("🔍 FAISS Recall Verification")
+    print("=" * 50)
+
+    # Check if files exist
+    if not Path(dataset_path).exists():
+        print(f"❌ Dataset not found: {dataset_path}")
+        return
+    if not Path(f"{index_path}.meta.json").exists():
+        print(f"❌ Index metadata not found: {index_path}.meta.json")
+        return
+
+    # Load data
+    print("📖 Loading FinanceBench queries...")
+    queries = load_financebench_queries(dataset_path, max_queries=50)
+    print(f"Loaded {len(queries)} queries")
+
+    print("📄 Loading passages from LEANN index...")
+    passages, passage_ids = load_passages_from_leann_index(index_path)
+
+    # Compute embeddings
+    print("🧮 Computing passage embeddings...")
+    passage_embeddings = compute_embeddings_direct(passages, embedding_model)
+
+    print("🧮 Computing query embeddings...")
+    query_embeddings = compute_embeddings_direct(queries, embedding_model)
+
+    # Build FAISS indexes
+    print("🏗️ Building FAISS indexes...")
+    flat_index, hnsw_index = build_faiss_indexes(passage_embeddings)
+
+    # Test different efSearch values (equivalent to LEANN complexity)
+    print("\n📊 Evaluating Recall@3 at different efSearch values...")
+    ef_search_values = [16, 32, 64, 128, 256]
+
+    for ef_search in ef_search_values:
+        print(f"\n🧪 Testing efSearch = {ef_search}")
+        start_time = time.time()
+
+        recall = evaluate_recall_at_k(
+            query_embeddings, flat_index, hnsw_index, passage_ids, k=3, ef_search=ef_search
+        )
+
+        elapsed = time.time() - start_time
+        print(
+            f"📈 efSearch {ef_search}: Recall@3 = {recall:.3f} ({recall * 100:.1f}%) in {elapsed:.2f}s"
+        )
+
+    print("\n✅ Verification completed!")
+    print("\n📋 Summary:")
+    print("  - Built independent FAISS Flat and HNSW indexes")
+    print("  - Compared recall@3 at different efSearch values")
+    print("  - Used same embedding model as LEANN")
+    print("  - This validates LEANN's recall measurements")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/laion/.gitignore

@@ -0,0 +1 @@
+data/

benchmarks/laion/README.md

@@ -0,0 +1,199 @@
+# LAION Multimodal Benchmark
+
+A multimodal benchmark for evaluating image retrieval and generation performance using LEANN with CLIP embeddings and Qwen2.5-VL for multimodal generation on LAION dataset subset.
+
+## Overview
+
+This benchmark evaluates:
+- **Image retrieval timing** using caption-based queries
+- **Recall@K performance** for image search
+- **Complexity analysis** across different search parameters
+- **Index size and storage efficiency**
+- **Multimodal generation** with Qwen2.5-VL for image understanding and description
+
+## Dataset Configuration
+
+- **Dataset**: LAION-400M subset (10,000 images)
+- **Embeddings**: Pre-computed CLIP ViT-B/32 (512 dimensions)
+- **Queries**: 200 random captions from the dataset
+- **Ground Truth**: Self-recall (query caption → original image)
+
+## Quick Start
+
+### 1. Setup the benchmark
+
+```bash
+cd benchmarks/laion
+python setup_laion.py --num-samples 10000 --num-queries 200
+```
+
+This will:
+- Create dummy LAION data (10K samples)
+- Generate CLIP embeddings (512-dim)
+- Build LEANN index with HNSW backend
+- Create 200 evaluation queries
+
+### 2. Run evaluation
+
+```bash
+# Run all evaluation stages
+python evaluate_laion.py --index data/laion_index.leann
+
+# Run specific stages
+python evaluate_laion.py --index data/laion_index.leann --stage 2  # Recall evaluation
+python evaluate_laion.py --index data/laion_index.leann --stage 3  # Complexity analysis
+python evaluate_laion.py --index data/laion_index.leann --stage 4  # Index comparison
+python evaluate_laion.py --index data/laion_index.leann --stage 5  # Multimodal generation
+
+# Multimodal generation with Qwen2.5-VL
+python evaluate_laion.py --index data/laion_index.leann --stage 5 --model-name Qwen/Qwen2.5-VL-7B-Instruct
+```
+
+### 3. Save results
+
+```bash
+python evaluate_laion.py --index data/laion_index.leann --output results.json
+```
+
+## Configuration Options
+
+### Setup Options
+```bash
+python setup_laion.py \
+  --num-samples 10000 \
+  --num-queries 200 \
+  --index-path data/laion_index.leann \
+  --backend hnsw
+```
+
+### Evaluation Options
+```bash
+python evaluate_laion.py \
+  --index data/laion_index.leann \
+  --queries data/evaluation_queries.jsonl \
+  --complexity 64 \
+  --top-k 3 \
+  --num-samples 100 \
+  --stage all
+```
+
+## Evaluation Stages
+
+### Stage 2: Recall Evaluation
+- Evaluates Recall@3 for multimodal retrieval
+- Compares LEANN vs FAISS baseline performance
+- Self-recall: query caption should retrieve original image
+
+### Stage 3: Complexity Analysis
+- Binary search for optimal complexity (90% recall target)
+- Tests performance across different complexity levels
+- Analyzes speed vs. accuracy tradeoffs
+
+### Stage 4: Index Comparison
+- Compares compact vs non-compact index sizes
+- Measures search performance differences
+- Reports storage efficiency and speed ratios
+
+### Stage 5: Multimodal Generation
+- Uses Qwen2.5-VL for image understanding and description
+- Retrieval-Augmented Generation (RAG) with multimodal context
+- Measures both search and generation timing
+
+## Output Metrics
+
+### Timing Metrics
+- Average/median/min/max search time
+- Standard deviation
+- Searches per second
+- Latency in milliseconds
+
+### Recall Metrics
+- Recall@3 percentage for image retrieval
+- Number of queries with ground truth
+
+### Index Metrics
+- Total index size (MB)
+- Component breakdown (index, passages, metadata)
+- Storage savings (compact vs non-compact)
+- Backend and embedding model info
+
+### Generation Metrics (Stage 5)
+- Average search time per query
+- Average generation time per query
+- Time distribution (search vs generation)
+- Sample multimodal responses
+- Model: Qwen2.5-VL performance
+
+## Benchmark Results
+
+### LEANN-RAG Performance (CLIP ViT-L/14 + Qwen2.5-VL)
+
+**Stage 3: Optimal Complexity Analysis**
+- **Optimal Complexity**: 85 (achieving 90% Recall@3)
+- **Binary Search Range**: 1-128
+- **Target Recall**: 90%
+- **Index Type**: Non-compact (for fast binary search)
+
+**Stage 5: Multimodal Generation Performance (Qwen2.5-VL)**
+- **Total Queries**: 20
+- **Average Search Time**: 1.200s per query
+- **Average Generation Time**: 6.558s per query
+- **Time Distribution**: Search 15.5%, Generation 84.5%
+- **LLM Backend**: HuggingFace transformers
+- **Model**: Qwen/Qwen2.5-VL-7B-Instruct
+- **Optimal Complexity**: 85
+
+**System Performance:**
+- **Index Size**: ~10,000 image embeddings from LAION subset
+- **Embedding Model**: CLIP ViT-L/14 (768 dimensions)
+- **Backend**: HNSW with cosine distance
+
+### Example Results
+
+```
+🎯 LAION MULTIMODAL BENCHMARK RESULTS
+============================================================
+
+📊 Multimodal Generation Results:
+  Total Queries: 20
+  Avg Search Time: 1.200s
+  Avg Generation Time: 6.558s
+  Time Distribution: Search 15.5%, Generation 84.5%
+  LLM Backend: HuggingFace transformers
+  Model: Qwen/Qwen2.5-VL-7B-Instruct
+
+⚙️ Optimal Complexity Analysis:
+  Target Recall: 90%
+  Optimal Complexity: 85
+  Binary Search Range: 1-128
+  Non-compact Index (fast search, no recompute)
+
+🚀 Performance Summary:
+  Multimodal RAG: 7.758s total per query
+  Search: 15.5% of total time
+  Generation: 84.5% of total time
+```
+
+## Directory Structure
+
+```
+benchmarks/laion/
+├── setup_laion.py           # Setup script
+├── evaluate_laion.py        # Evaluation script
+├── README.md               # This file
+└── data/                   # Generated data
+    ├── laion_images/       # Image files (placeholder)
+    ├── laion_metadata.jsonl # Image metadata
+    ├── laion_passages.jsonl # LEANN passages
+    ├── laion_embeddings.npy # CLIP embeddings
+    ├── evaluation_queries.jsonl # Evaluation queries
+    └── laion_index.leann/  # LEANN index files
+```
+
+## Notes
+
+- Current implementation uses dummy data for demonstration
+- For real LAION data, implement actual download logic in `setup_laion.py`
+- CLIP embeddings are randomly generated - replace with real CLIP model for production
+- Adjust `num_samples` and `num_queries` based on available resources
+- Consider using `--num-samples` during evaluation for faster testing

benchmarks/laion/evaluate_laion.py

@@ -0,0 +1,725 @@
+"""
+LAION Multimodal Benchmark Evaluation Script - Modular Recall-based Evaluation
+"""
+
+import argparse
+import json
+import logging
+import os
+import pickle
+import time
+from pathlib import Path
+
+import numpy as np
+from leann import LeannSearcher
+from leann_backend_hnsw import faiss
+from sentence_transformers import SentenceTransformer
+
+from ..llm_utils import evaluate_multimodal_rag, load_qwen_vl_model
+
+# Setup logging to reduce verbose output
+logging.basicConfig(level=logging.WARNING)
+logging.getLogger("leann.api").setLevel(logging.WARNING)
+logging.getLogger("leann_backend_hnsw").setLevel(logging.WARNING)
+
+
+class RecallEvaluator:
+    """Stage 2: Evaluate Recall@3 (LEANN vs FAISS baseline for multimodal retrieval)"""
+
+    def __init__(self, index_path: str, baseline_dir: str):
+        self.index_path = index_path
+        self.baseline_dir = baseline_dir
+        self.searcher = LeannSearcher(index_path)
+
+        # Load FAISS flat baseline (image embeddings)
+        baseline_index_path = os.path.join(baseline_dir, "faiss_flat.index")
+        metadata_path = os.path.join(baseline_dir, "metadata.pkl")
+
+        self.faiss_index = faiss.read_index(baseline_index_path)
+        with open(metadata_path, "rb") as f:
+            self.image_ids = pickle.load(f)
+        print(f"📚 Loaded FAISS flat baseline with {self.faiss_index.ntotal} image vectors")
+
+        # Load sentence-transformers CLIP for text embedding (ViT-L/14)
+        self.st_clip = SentenceTransformer("clip-ViT-L-14")
+
+    def evaluate_recall_at_3(
+        self, captions: list[str], complexity: int = 64, recompute_embeddings: bool = True
+    ) -> float:
+        """Evaluate recall@3 for multimodal retrieval: caption queries -> image results"""
+        recompute_str = "with recompute" if recompute_embeddings else "no recompute"
+        print(f"🔍 Evaluating recall@3 with complexity={complexity} ({recompute_str})...")
+
+        total_recall = 0.0
+        num_queries = len(captions)
+
+        for i, caption in enumerate(captions):
+            # Get ground truth: search with FAISS flat using caption text embedding
+            # Generate CLIP text embedding for caption via sentence-transformers (normalized)
+            query_embedding = self.st_clip.encode(
+                [caption], convert_to_numpy=True, normalize_embeddings=True, show_progress_bar=False
+            ).astype(np.float32)
+
+            # Search FAISS flat for ground truth using LEANN's modified faiss API
+            n = query_embedding.shape[0]  # Number of queries
+            k = 3  # Number of nearest neighbors
+            distances = np.zeros((n, k), dtype=np.float32)
+            labels = np.zeros((n, k), dtype=np.int64)
+
+            self.faiss_index.search(
+                n,
+                faiss.swig_ptr(query_embedding),
+                k,
+                faiss.swig_ptr(distances),
+                faiss.swig_ptr(labels),
+            )
+
+            # Extract the results (image IDs from FAISS)
+            baseline_ids = {self.image_ids[idx] for idx in labels[0]}
+
+            # Search with LEANN at specified complexity (using caption as text query)
+            test_results = self.searcher.search(
+                caption,
+                top_k=3,
+                complexity=complexity,
+                recompute_embeddings=recompute_embeddings,
+            )
+            test_ids = {result.id for result in test_results}
+
+            # Calculate recall@3 = |intersection| / |ground_truth|
+            intersection = test_ids.intersection(baseline_ids)
+            recall = len(intersection) / 3.0  # Ground truth size is 3
+            total_recall += recall
+
+            if i < 3:  # Show first few examples
+                print(f"  Query {i + 1}: '{caption[:50]}...' -> Recall@3: {recall:.3f}")
+                print(f"    FAISS ground truth: {list(baseline_ids)}")
+                print(f"    LEANN results (C={complexity}, {recompute_str}): {list(test_ids)}")
+                print(f"    Intersection: {list(intersection)}")
+
+        avg_recall = total_recall / num_queries
+        print(f"📊 Average Recall@3: {avg_recall:.3f} ({avg_recall * 100:.1f}%)")
+        return avg_recall
+
+    def cleanup(self):
+        """Cleanup resources"""
+        if hasattr(self, "searcher"):
+            self.searcher.cleanup()
+
+
+class LAIONEvaluator:
+    def __init__(self, index_path: str):
+        self.index_path = index_path
+        self.searcher = LeannSearcher(index_path)
+
+    def load_queries(self, queries_file: str) -> list[str]:
+        """Load caption queries from evaluation file"""
+        captions = []
+        with open(queries_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    query_data = json.loads(line)
+                    captions.append(query_data["query"])
+
+        print(f"📊 Loaded {len(captions)} caption queries")
+        return captions
+
+    def analyze_index_sizes(self) -> dict:
+        """Analyze index sizes, emphasizing .index only (exclude passages)."""
+        print("📏 Analyzing index sizes (.index only)...")
+
+        # Get all index-related files
+        index_path = Path(self.index_path)
+        index_dir = index_path.parent
+        index_name = index_path.stem  # Remove .leann extension
+
+        sizes: dict[str, float] = {}
+
+        # Core index files
+        index_file = index_dir / f"{index_name}.index"
+        meta_file = index_dir / f"{index_path.name}.meta.json"  # Keep .leann for meta file
+        passages_file = index_dir / f"{index_path.name}.passages.jsonl"  # Keep .leann for passages
+        passages_idx_file = index_dir / f"{index_path.name}.passages.idx"  # Keep .leann for idx
+
+        # Core index size (.index only)
+        index_mb = index_file.stat().st_size / (1024 * 1024) if index_file.exists() else 0.0
+        sizes["index_only_mb"] = index_mb
+
+        # Other files for reference (not counted in index_only_mb)
+        sizes["metadata_mb"] = (
+            meta_file.stat().st_size / (1024 * 1024) if meta_file.exists() else 0.0
+        )
+        sizes["passages_text_mb"] = (
+            passages_file.stat().st_size / (1024 * 1024) if passages_file.exists() else 0.0
+        )
+        sizes["passages_index_mb"] = (
+            passages_idx_file.stat().st_size / (1024 * 1024) if passages_idx_file.exists() else 0.0
+        )
+
+        print(f"  📁 .index size: {index_mb:.1f} MB")
+        if sizes["metadata_mb"]:
+            print(f"  🧾 metadata: {sizes['metadata_mb']:.3f} MB")
+        if sizes["passages_text_mb"] or sizes["passages_index_mb"]:
+            print(
+                f"  (passages excluded) text: {sizes['passages_text_mb']:.1f} MB, idx: {sizes['passages_index_mb']:.1f} MB"
+            )
+
+        return sizes
+
+    def create_non_compact_index_for_comparison(self, non_compact_index_path: str) -> dict:
+        """Create a non-compact index for comparison purposes"""
+        print("🏗️ Building non-compact index from existing passages...")
+
+        # Load existing passages from current index
+        from leann import LeannBuilder
+
+        current_index_path = Path(self.index_path)
+        current_index_dir = current_index_path.parent
+        current_index_name = current_index_path.name
+
+        # Read metadata to get passage source
+        meta_path = current_index_dir / f"{current_index_name}.meta.json"
+        with open(meta_path) as f:
+            meta = json.load(f)
+
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not Path(passage_file).is_absolute():
+            passage_file = current_index_dir / Path(passage_file).name
+
+        print(f"📄 Loading passages from {passage_file}...")
+
+        # Load CLIP embeddings
+        embeddings_file = current_index_dir / "clip_image_embeddings.npy"
+        embeddings = np.load(embeddings_file)
+        print(f"📐 Loaded embeddings shape: {embeddings.shape}")
+
+        # Build non-compact index with same passages and embeddings
+        builder = LeannBuilder(
+            backend_name="hnsw",
+            # Use CLIP text encoder (ViT-L/14) to match image embeddings (768-dim)
+            embedding_model="clip-ViT-L-14",
+            embedding_mode="sentence-transformers",
+            is_recompute=False,  # Disable recompute (store embeddings)
+            is_compact=False,  # Disable compact storage
+            distance_metric="cosine",
+            **{
+                k: v
+                for k, v in meta.get("backend_kwargs", {}).items()
+                if k not in ["is_recompute", "is_compact", "distance_metric"]
+            },
+        )
+
+        # Prepare ids and add passages
+        ids: list[str] = []
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    ids.append(str(data["id"]))
+                    # Ensure metadata contains the id used by the vector index
+                    metadata = {**data.get("metadata", {}), "id": data["id"]}
+                    builder.add_text(text=data["text"], metadata=metadata)
+
+        if len(ids) != embeddings.shape[0]:
+            raise ValueError(
+                f"IDs count ({len(ids)}) does not match embeddings ({embeddings.shape[0]})."
+            )
+
+        # Persist a pickle for build_index_from_embeddings
+        pkl_path = current_index_dir / "clip_image_embeddings.pkl"
+        with open(pkl_path, "wb") as pf:
+            pickle.dump((ids, embeddings.astype(np.float32)), pf)
+
+        print(
+            f"🔨 Building non-compact index at {non_compact_index_path} from precomputed embeddings..."
+        )
+        builder.build_index_from_embeddings(non_compact_index_path, str(pkl_path))
+
+        # Analyze the non-compact index size
+        temp_evaluator = LAIONEvaluator(non_compact_index_path)
+        non_compact_sizes = temp_evaluator.analyze_index_sizes()
+        non_compact_sizes["index_type"] = "non_compact"
+
+        return non_compact_sizes
+
+    def compare_index_performance(
+        self, non_compact_path: str, compact_path: str, test_captions: list, complexity: int
+    ) -> dict:
+        """Compare performance between non-compact and compact indexes"""
+        print("⚡ Comparing search performance between indexes...")
+
+        # Test queries
+        test_queries = test_captions[:5]
+
+        results = {
+            "non_compact": {"search_times": []},
+            "compact": {"search_times": []},
+            "avg_search_times": {},
+            "speed_ratio": 0.0,
+        }
+
+        # Test non-compact index (no recompute)
+        print("  🔍 Testing non-compact index (no recompute)...")
+        non_compact_searcher = LeannSearcher(non_compact_path)
+
+        for caption in test_queries:
+            start_time = time.time()
+            _ = non_compact_searcher.search(
+                caption, top_k=3, complexity=complexity, recompute_embeddings=False
+            )
+            search_time = time.time() - start_time
+            results["non_compact"]["search_times"].append(search_time)
+
+        # Test compact index (with recompute)
+        print("  🔍 Testing compact index (with recompute)...")
+        compact_searcher = LeannSearcher(compact_path)
+
+        for caption in test_queries:
+            start_time = time.time()
+            _ = compact_searcher.search(
+                caption, top_k=3, complexity=complexity, recompute_embeddings=True
+            )
+            search_time = time.time() - start_time
+            results["compact"]["search_times"].append(search_time)
+
+        # Calculate averages
+        results["avg_search_times"]["non_compact"] = sum(
+            results["non_compact"]["search_times"]
+        ) / len(results["non_compact"]["search_times"])
+        results["avg_search_times"]["compact"] = sum(results["compact"]["search_times"]) / len(
+            results["compact"]["search_times"]
+        )
+
+        # Performance ratio
+        if results["avg_search_times"]["compact"] > 0:
+            results["speed_ratio"] = (
+                results["avg_search_times"]["non_compact"] / results["avg_search_times"]["compact"]
+            )
+        else:
+            results["speed_ratio"] = float("inf")
+
+        print(
+            f"    Non-compact (no recompute): {results['avg_search_times']['non_compact']:.3f}s avg"
+        )
+        print(f"    Compact (with recompute): {results['avg_search_times']['compact']:.3f}s avg")
+        print(f"    Speed ratio: {results['speed_ratio']:.2f}x")
+
+        # Cleanup
+        non_compact_searcher.cleanup()
+        compact_searcher.cleanup()
+
+        return results
+
+    def _print_results(self, timing_metrics: dict):
+        """Print evaluation results"""
+        print("\n🎯 LAION MULTIMODAL BENCHMARK RESULTS")
+        print("=" * 60)
+
+        # Index comparison analysis (prefer .index-only view if present)
+        if "current_index" in timing_metrics and "non_compact_index" in timing_metrics:
+            current = timing_metrics["current_index"]
+            non_compact = timing_metrics["non_compact_index"]
+
+            if "index_only_mb" in current and "index_only_mb" in non_compact:
+                print("\n📏 Index Comparison Analysis (.index only):")
+                print(f"  Compact index (current): {current.get('index_only_mb', 0):.1f} MB")
+                print(f"  Non-compact index: {non_compact.get('index_only_mb', 0):.1f} MB")
+                print(
+                    f"  Storage saving by compact: {timing_metrics.get('storage_saving_percent', 0):.1f}%"
+                )
+                # Show excluded components for reference if available
+                if any(
+                    k in non_compact
+                    for k in ("passages_text_mb", "passages_index_mb", "metadata_mb")
+                ):
+                    print("  (passages excluded in totals, shown for reference):")
+                    print(
+                        f"    - Passages text: {non_compact.get('passages_text_mb', 0):.1f} MB, "
+                        f"Passages index: {non_compact.get('passages_index_mb', 0):.1f} MB, "
+                        f"Metadata: {non_compact.get('metadata_mb', 0):.3f} MB"
+                    )
+            else:
+                # Fallback to legacy totals if running with older metrics
+                print("\n📏 Index Comparison Analysis:")
+                print(
+                    f"  Compact index (current): {current.get('total_with_embeddings', 0):.1f} MB"
+                )
+                print(
+                    f"  Non-compact index (with embeddings): {non_compact.get('total_with_embeddings', 0):.1f} MB"
+                )
+                print(
+                    f"  Storage saving by compact: {timing_metrics.get('storage_saving_percent', 0):.1f}%"
+                )
+                print("  Component breakdown (non-compact):")
+                print(f"    - Main index: {non_compact.get('index', 0):.1f} MB")
+                print(f"    - Passages text: {non_compact.get('passages_text', 0):.1f} MB")
+                print(f"    - Passages index: {non_compact.get('passages_index', 0):.1f} MB")
+                print(f"    - Metadata: {non_compact.get('metadata', 0):.1f} MB")
+
+        # Performance comparison
+        if "performance_comparison" in timing_metrics:
+            perf = timing_metrics["performance_comparison"]
+            print("\n⚡ Performance Comparison:")
+            print(
+                f"  Non-compact (no recompute): {perf.get('avg_search_times', {}).get('non_compact', 0):.3f}s avg"
+            )
+            print(
+                f"  Compact (with recompute): {perf.get('avg_search_times', {}).get('compact', 0):.3f}s avg"
+            )
+            print(f"  Speed ratio: {perf.get('speed_ratio', 0):.2f}x")
+
+        # Legacy single index analysis (fallback)
+        if "total_with_embeddings" in timing_metrics and "current_index" not in timing_metrics:
+            print("\n📏 Index Size Analysis:")
+            print(
+                f"  Index with embeddings: {timing_metrics.get('total_with_embeddings', 0):.1f} MB"
+            )
+            print(
+                f"  Estimated pruned index: {timing_metrics.get('total_without_embeddings', 0):.1f} MB"
+            )
+            print(f"  Compression ratio: {timing_metrics.get('compression_ratio', 0):.2f}x")
+
+    def cleanup(self):
+        """Cleanup resources"""
+        if self.searcher:
+            self.searcher.cleanup()
+
+
+def main():
+    parser = argparse.ArgumentParser(description="LAION Multimodal Benchmark Evaluation")
+    parser.add_argument("--index", required=True, help="Path to LEANN index")
+    parser.add_argument(
+        "--queries", default="data/evaluation_queries.jsonl", help="Path to evaluation queries"
+    )
+    parser.add_argument(
+        "--stage",
+        choices=["2", "3", "4", "5", "all"],
+        default="all",
+        help="Which stage to run (2=recall, 3=complexity, 4=index comparison, 5=generation)",
+    )
+    parser.add_argument("--complexity", type=int, default=None, help="Complexity for search")
+    parser.add_argument("--baseline-dir", default="baseline", help="Baseline output directory")
+    parser.add_argument("--output", help="Save results to JSON file")
+    parser.add_argument(
+        "--llm-backend",
+        choices=["hf"],
+        default="hf",
+        help="LLM backend (Qwen2.5-VL only supports HF)",
+    )
+    parser.add_argument(
+        "--model-name", default="Qwen/Qwen2.5-VL-7B-Instruct", help="Multimodal model name"
+    )
+
+    args = parser.parse_args()
+
+    try:
+        # Check if baseline exists
+        baseline_index_path = os.path.join(args.baseline_dir, "faiss_flat.index")
+        if not os.path.exists(baseline_index_path):
+            print(f"❌ FAISS baseline not found at {baseline_index_path}")
+            print("💡 Please run setup_laion.py first to build the baseline")
+            exit(1)
+
+        if args.stage == "2" or args.stage == "all":
+            # Stage 2: Recall@3 evaluation
+            print("🚀 Starting Stage 2: Recall@3 evaluation for multimodal retrieval")
+
+            evaluator = RecallEvaluator(args.index, args.baseline_dir)
+
+            # Load caption queries for testing
+            laion_evaluator = LAIONEvaluator(args.index)
+            captions = laion_evaluator.load_queries(args.queries)
+
+            # Test with queries for robust measurement
+            test_captions = captions[:100]  # Use subset for speed
+            print(f"🧪 Testing with {len(test_captions)} caption queries")
+
+            # Test with complexity 64
+            complexity = 64
+            recall = evaluator.evaluate_recall_at_3(test_captions, complexity)
+            print(f"📈 Recall@3 at complexity {complexity}: {recall * 100:.1f}%")
+
+            evaluator.cleanup()
+            print("✅ Stage 2 completed!\n")
+
+        # Shared non-compact index path for Stage 3 and 4
+        non_compact_index_path = args.index.replace(".leann", "_noncompact.leann")
+        complexity = args.complexity
+
+        if args.stage == "3" or args.stage == "all":
+            # Stage 3: Binary search for 90% recall complexity
+            print("🚀 Starting Stage 3: Binary search for 90% recall complexity")
+            print(
+                "💡 Creating non-compact index for fast binary search with recompute_embeddings=False"
+            )
+
+            # Create non-compact index for binary search
+            print("🏗️ Creating non-compact index for binary search...")
+            evaluator = LAIONEvaluator(args.index)
+            evaluator.create_non_compact_index_for_comparison(non_compact_index_path)
+
+            # Use non-compact index for binary search
+            binary_search_evaluator = RecallEvaluator(non_compact_index_path, args.baseline_dir)
+
+            # Load caption queries for testing
+            captions = evaluator.load_queries(args.queries)
+
+            # Use subset for robust measurement
+            test_captions = captions[:50]  # Smaller subset for binary search speed
+            print(f"🧪 Testing with {len(test_captions)} caption queries")
+
+            # Binary search for 90% recall complexity
+            target_recall = 0.9
+            min_complexity, max_complexity = 1, 128
+
+            print(f"🔍 Binary search for {target_recall * 100}% recall complexity...")
+            print(f"Search range: {min_complexity} to {max_complexity}")
+
+            best_complexity = None
+            best_recall = 0.0
+
+            while min_complexity <= max_complexity:
+                mid_complexity = (min_complexity + max_complexity) // 2
+
+                print(
+                    f"\n🧪 Testing complexity {mid_complexity} (no recompute, non-compact index)..."
+                )
+                # Use recompute_embeddings=False on non-compact index for fast binary search
+                recall = binary_search_evaluator.evaluate_recall_at_3(
+                    test_captions, mid_complexity, recompute_embeddings=False
+                )
+
+                print(
+                    f"  Complexity {mid_complexity}: Recall@3 = {recall:.3f} ({recall * 100:.1f}%)"
+                )
+
+                if recall >= target_recall:
+                    best_complexity = mid_complexity
+                    best_recall = recall
+                    max_complexity = mid_complexity - 1
+                    print("  ✅ Target reached! Searching for lower complexity...")
+                else:
+                    min_complexity = mid_complexity + 1
+                    print("  ❌ Below target. Searching for higher complexity...")
+
+            if best_complexity is not None:
+                print("\n🎯 Optimal complexity found!")
+                print(f"  Complexity: {best_complexity}")
+                print(f"  Recall@3: {best_recall:.3f} ({best_recall * 100:.1f}%)")
+
+                # Test a few complexities around the optimal one for verification
+                print("\n🔬 Verification test around optimal complexity:")
+                verification_complexities = [
+                    max(1, best_complexity - 2),
+                    max(1, best_complexity - 1),
+                    best_complexity,
+                    best_complexity + 1,
+                    best_complexity + 2,
+                ]
+
+                for complexity in verification_complexities:
+                    if complexity <= 512:  # reasonable upper bound
+                        recall = binary_search_evaluator.evaluate_recall_at_3(
+                            test_captions, complexity, recompute_embeddings=False
+                        )
+                        status = "✅" if recall >= target_recall else "❌"
+                        print(f"  {status} Complexity {complexity:3d}: {recall * 100:5.1f}%")
+
+                # Now test the optimal complexity with compact index and recompute for comparison
+                print(
+                    f"\n🔄 Testing optimal complexity {best_complexity} on compact index WITH recompute..."
+                )
+                compact_evaluator = RecallEvaluator(args.index, args.baseline_dir)
+                recall_with_recompute = compact_evaluator.evaluate_recall_at_3(
+                    test_captions[:10], best_complexity, recompute_embeddings=True
+                )
+                print(
+                    f"  ✅ Complexity {best_complexity} (compact index with recompute): {recall_with_recompute * 100:.1f}%"
+                )
+                complexity = best_complexity
+                print(
+                    f"  📊 Recall difference: {abs(best_recall - recall_with_recompute) * 100:.2f}%"
+                )
+                compact_evaluator.cleanup()
+            else:
+                print(f"\n❌ Could not find complexity achieving {target_recall * 100}% recall")
+                print("All tested complexities were below target.")
+
+            # Cleanup evaluators (keep non-compact index for Stage 4)
+            binary_search_evaluator.cleanup()
+            evaluator.cleanup()
+
+            print("✅ Stage 3 completed! Non-compact index saved for Stage 4.\n")
+
+        if args.stage == "4" or args.stage == "all":
+            # Stage 4: Index comparison (without LLM generation)
+            print("🚀 Starting Stage 4: Index comparison analysis")
+
+            # Use LAION evaluator for index comparison
+            evaluator = LAIONEvaluator(args.index)
+
+            # Load caption queries
+            captions = evaluator.load_queries(args.queries)
+
+            # Step 1: Analyze current (compact) index
+            print("\n📏 Analyzing current index (compact, pruned)...")
+            compact_size_metrics = evaluator.analyze_index_sizes()
+            compact_size_metrics["index_type"] = "compact"
+
+            # Step 2: Use existing non-compact index or create if needed
+            if Path(non_compact_index_path).exists():
+                print(
+                    f"\n📁 Using existing non-compact index from Stage 3: {non_compact_index_path}"
+                )
+                temp_evaluator = LAIONEvaluator(non_compact_index_path)
+                non_compact_size_metrics = temp_evaluator.analyze_index_sizes()
+                non_compact_size_metrics["index_type"] = "non_compact"
+            else:
+                print("\n🏗️ Creating non-compact index (with embeddings) for comparison...")
+                non_compact_size_metrics = evaluator.create_non_compact_index_for_comparison(
+                    non_compact_index_path
+                )
+
+            # Step 3: Compare index sizes (.index only)
+            print("\n📊 Index size comparison (.index only):")
+            print(
+                f"  Compact index (current): {compact_size_metrics.get('index_only_mb', 0):.1f} MB"
+            )
+            print(f"  Non-compact index: {non_compact_size_metrics.get('index_only_mb', 0):.1f} MB")
+
+            storage_saving = 0.0
+            if non_compact_size_metrics.get("index_only_mb", 0) > 0:
+                storage_saving = (
+                    (
+                        non_compact_size_metrics.get("index_only_mb", 0)
+                        - compact_size_metrics.get("index_only_mb", 0)
+                    )
+                    / non_compact_size_metrics.get("index_only_mb", 1)
+                    * 100
+                )
+            print(f"  Storage saving by compact: {storage_saving:.1f}%")
+
+            # Step 4: Performance comparison between the two indexes
+            if complexity is None:
+                raise ValueError("Complexity is required for index comparison")
+
+            print("\n⚡ Performance comparison between indexes...")
+            performance_metrics = evaluator.compare_index_performance(
+                non_compact_index_path, args.index, captions[:10], complexity=complexity
+            )
+
+            # Combine all metrics
+            combined_metrics = {
+                "current_index": compact_size_metrics,
+                "non_compact_index": non_compact_size_metrics,
+                "performance_comparison": performance_metrics,
+                "storage_saving_percent": storage_saving,
+            }
+
+            # Print comprehensive results
+            evaluator._print_results(combined_metrics)
+
+            # Save results if requested
+            if args.output:
+                print(f"\n💾 Saving results to {args.output}...")
+                with open(args.output, "w") as f:
+                    json.dump(combined_metrics, f, indent=2, default=str)
+                print(f"✅ Results saved to {args.output}")
+
+            evaluator.cleanup()
+            print("✅ Stage 4 completed!\n")
+
+        if args.stage in ("5", "all"):
+            print("🚀 Starting Stage 5: Multimodal generation with Qwen2.5-VL")
+            evaluator = LAIONEvaluator(args.index)
+            captions = evaluator.load_queries(args.queries)
+            test_captions = captions[: min(20, len(captions))]  # Use subset for generation
+
+            print(f"🧪 Testing multimodal generation with {len(test_captions)} queries")
+
+            # Load Qwen2.5-VL model
+            try:
+                print("Loading Qwen2.5-VL model...")
+                processor, model = load_qwen_vl_model(args.model_name)
+
+                # Run multimodal generation evaluation
+                complexity = args.complexity or 64
+                gen_results = evaluate_multimodal_rag(
+                    evaluator.searcher,
+                    test_captions,
+                    processor=processor,
+                    model=model,
+                    complexity=complexity,
+                )
+
+                print("\n📊 Multimodal Generation Results:")
+                print(f"  Total Queries: {len(test_captions)}")
+                print(f"  Avg Search Time: {gen_results['avg_search_time']:.3f}s")
+                print(f"  Avg Generation Time: {gen_results['avg_generation_time']:.3f}s")
+                total_time = gen_results["avg_search_time"] + gen_results["avg_generation_time"]
+                search_pct = (gen_results["avg_search_time"] / total_time) * 100
+                gen_pct = (gen_results["avg_generation_time"] / total_time) * 100
+                print(f"  Time Distribution: Search {search_pct:.1f}%, Generation {gen_pct:.1f}%")
+                print("  LLM Backend: HuggingFace transformers")
+                print(f"  Model: {args.model_name}")
+
+                # Show sample results
+                print("\n📝 Sample Multimodal Generations:")
+                for i, response in enumerate(gen_results["results"][:3]):
+                    # Handle both string and dict formats for captions
+                    if isinstance(test_captions[i], dict):
+                        caption_text = test_captions[i].get("query", str(test_captions[i]))
+                    else:
+                        caption_text = str(test_captions[i])
+                    print(f"  Query {i + 1}: {caption_text[:60]}...")
+                    print(f"  Response {i + 1}: {response[:100]}...")
+                    print()
+
+            except Exception as e:
+                print(f"❌ Multimodal generation evaluation failed: {e}")
+                print("💡 Make sure transformers and Qwen2.5-VL are installed")
+                import traceback
+
+                traceback.print_exc()
+
+            evaluator.cleanup()
+            print("✅ Stage 5 completed!\n")
+
+        if args.stage == "all":
+            print("🎉 All evaluation stages completed successfully!")
+            print("\n📋 Summary:")
+            print("  Stage 2: ✅ Multimodal Recall@3 evaluation completed")
+            print("  Stage 3: ✅ Optimal complexity found")
+            print("  Stage 4: ✅ Index comparison analysis completed")
+            print("  Stage 5: ✅ Multimodal generation evaluation completed")
+            print("\n🔧 Recommended next steps:")
+            print("  - Use optimal complexity for best speed/accuracy balance")
+            print("  - Review index comparison for storage vs performance tradeoffs")
+
+            # Clean up non-compact index after all stages complete
+            print("\n🧹 Cleaning up temporary non-compact index...")
+            if Path(non_compact_index_path).exists():
+                temp_index_dir = Path(non_compact_index_path).parent
+                temp_index_name = Path(non_compact_index_path).name
+                for temp_file in temp_index_dir.glob(f"{temp_index_name}*"):
+                    temp_file.unlink()
+                print(f"✅ Cleaned up {non_compact_index_path}")
+            else:
+                print("📝 No temporary index to clean up")
+
+    except KeyboardInterrupt:
+        print("\n⚠️  Evaluation interrupted by user")
+        exit(1)
+    except Exception as e:
+        print(f"\n❌ Stage {args.stage} failed: {e}")
+        import traceback
+
+        traceback.print_exc()
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/laion/setup_laion.py

@@ -0,0 +1,576 @@
+"""
+LAION Multimodal Benchmark Setup Script
+Downloads LAION subset and builds LEANN index with sentence embeddings
+"""
+
+import argparse
+import asyncio
+import io
+import json
+import os
+import pickle
+import time
+from pathlib import Path
+
+import aiohttp
+import numpy as np
+from datasets import load_dataset
+from leann import LeannBuilder
+from PIL import Image
+from sentence_transformers import SentenceTransformer
+from tqdm import tqdm
+
+
+class LAIONSetup:
+    def __init__(self, data_dir: str = "data"):
+        self.data_dir = Path(data_dir)
+        self.images_dir = self.data_dir / "laion_images"
+        self.metadata_file = self.data_dir / "laion_metadata.jsonl"
+
+        # Create directories
+        self.data_dir.mkdir(exist_ok=True)
+        self.images_dir.mkdir(exist_ok=True)
+
+    async def download_single_image(self, session, sample_data, semaphore, progress_bar):
+        """Download a single image asynchronously"""
+        async with semaphore:  # Limit concurrent downloads
+            try:
+                image_url = sample_data["url"]
+                image_path = sample_data["image_path"]
+
+                # Skip if already exists
+                if os.path.exists(image_path):
+                    progress_bar.update(1)
+                    return sample_data
+
+                async with session.get(image_url, timeout=10) as response:
+                    if response.status == 200:
+                        content = await response.read()
+
+                        # Verify it's a valid image
+                        try:
+                            img = Image.open(io.BytesIO(content))
+                            img = img.convert("RGB")
+                            img.save(image_path, "JPEG")
+                            progress_bar.update(1)
+                            return sample_data
+                        except Exception:
+                            progress_bar.update(1)
+                            return None  # Skip invalid images
+                    else:
+                        progress_bar.update(1)
+                        return None
+
+            except Exception:
+                progress_bar.update(1)
+                return None
+
+    def download_laion_subset(self, num_samples: int = 1000):
+        """Download LAION subset from HuggingFace datasets with async parallel downloading"""
+        print(f"📥 Downloading LAION subset ({num_samples} samples)...")
+
+        # Load LAION-400M subset from HuggingFace
+        print("🤗 Loading from HuggingFace datasets...")
+        dataset = load_dataset("laion/laion400m", split="train", streaming=True)
+
+        # Collect sample metadata first (fast)
+        print("📋 Collecting sample metadata...")
+        candidates = []
+        for sample in dataset:
+            if len(candidates) >= num_samples * 3:  # Get 3x more candidates in case some fail
+                break
+
+            image_url = sample.get("url", "")
+            caption = sample.get("caption", "")
+
+            if not image_url or not caption:
+                continue
+
+            image_filename = f"laion_{len(candidates):06d}.jpg"
+            image_path = self.images_dir / image_filename
+
+            candidate = {
+                "id": f"laion_{len(candidates):06d}",
+                "url": image_url,
+                "caption": caption,
+                "image_path": str(image_path),
+                "width": sample.get("original_width", 512),
+                "height": sample.get("original_height", 512),
+                "similarity": sample.get("similarity", 0.0),
+            }
+            candidates.append(candidate)
+
+        print(
+            f"📊 Collected {len(candidates)} candidates, downloading {num_samples} in parallel..."
+        )
+
+        # Download images in parallel
+        async def download_batch():
+            semaphore = asyncio.Semaphore(20)  # Limit to 20 concurrent downloads
+            connector = aiohttp.TCPConnector(limit=100, limit_per_host=20)
+            timeout = aiohttp.ClientTimeout(total=30)
+
+            progress_bar = tqdm(total=len(candidates[: num_samples * 2]), desc="Downloading images")
+
+            async with aiohttp.ClientSession(connector=connector, timeout=timeout) as session:
+                tasks = []
+                for candidate in candidates[: num_samples * 2]:  # Try 2x more than needed
+                    task = self.download_single_image(session, candidate, semaphore, progress_bar)
+                    tasks.append(task)
+
+                # Wait for all downloads
+                results = await asyncio.gather(*tasks, return_exceptions=True)
+                progress_bar.close()
+
+                # Filter successful downloads
+                successful = [r for r in results if r is not None and not isinstance(r, Exception)]
+                return successful[:num_samples]
+
+        # Run async download
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+        try:
+            samples = loop.run_until_complete(download_batch())
+        finally:
+            loop.close()
+
+        # Save metadata
+        with open(self.metadata_file, "w", encoding="utf-8") as f:
+            for sample in samples:
+                f.write(json.dumps(sample) + "\n")
+
+        print(f"✅ Downloaded {len(samples)} real LAION samples with async parallel downloading")
+        return samples
+
+    def generate_clip_image_embeddings(self, samples: list[dict]):
+        """Generate CLIP image embeddings for downloaded images"""
+        print("🔍 Generating CLIP image embeddings...")
+
+        # Load sentence-transformers CLIP (ViT-L/14, 768-dim) for image embeddings
+        # This single model can encode both images and text.
+        model = SentenceTransformer("clip-ViT-L-14")
+
+        embeddings = []
+        valid_samples = []
+
+        for sample in tqdm(samples, desc="Processing images"):
+            try:
+                # Load image
+                image_path = sample["image_path"]
+                image = Image.open(image_path).convert("RGB")
+
+                # Encode image to 768-dim embedding via sentence-transformers (normalized)
+                vec = model.encode(
+                    [image],
+                    convert_to_numpy=True,
+                    normalize_embeddings=True,
+                    batch_size=1,
+                    show_progress_bar=False,
+                )[0]
+                embeddings.append(vec.astype(np.float32))
+                valid_samples.append(sample)
+
+            except Exception as e:
+                print(f"  ⚠️ Failed to process {sample['id']}: {e}")
+                # Skip invalid images
+
+        embeddings = np.array(embeddings, dtype=np.float32)
+
+        # Save embeddings
+        embeddings_file = self.data_dir / "clip_image_embeddings.npy"
+        np.save(embeddings_file, embeddings)
+        print(f"✅ Generated {len(embeddings)} image embeddings, shape: {embeddings.shape}")
+
+        return embeddings, valid_samples
+
+    def build_faiss_baseline(
+        self, embeddings: np.ndarray, samples: list[dict], output_dir: str = "baseline"
+    ):
+        """Build FAISS flat baseline using CLIP image embeddings"""
+        print("🔨 Building FAISS Flat baseline...")
+
+        from leann_backend_hnsw import faiss
+
+        os.makedirs(output_dir, exist_ok=True)
+        baseline_path = os.path.join(output_dir, "faiss_flat.index")
+        metadata_path = os.path.join(output_dir, "metadata.pkl")
+
+        if os.path.exists(baseline_path) and os.path.exists(metadata_path):
+            print(f"✅ Baseline already exists at {baseline_path}")
+            return baseline_path
+
+        # Extract image IDs (must be present)
+        if not samples or "id" not in samples[0]:
+            raise KeyError("samples missing 'id' field for FAISS baseline")
+        image_ids: list[str] = [str(sample["id"]) for sample in samples]
+
+        print(f"📐 Embedding shape: {embeddings.shape}")
+        print(f"📄 Processing {len(image_ids)} images")
+
+        # Build FAISS flat index
+        print("🏗️ Building FAISS IndexFlatIP...")
+        dimension = embeddings.shape[1]
+        index = faiss.IndexFlatIP(dimension)
+
+        # Add embeddings to flat index
+        embeddings_f32 = embeddings.astype(np.float32)
+        index.add(embeddings_f32.shape[0], faiss.swig_ptr(embeddings_f32))
+
+        # Save index and metadata
+        faiss.write_index(index, baseline_path)
+        with open(metadata_path, "wb") as f:
+            pickle.dump(image_ids, f)
+
+        print(f"✅ FAISS baseline saved to {baseline_path}")
+        print(f"✅ Metadata saved to {metadata_path}")
+        print(f"📊 Total vectors: {index.ntotal}")
+
+        return baseline_path
+
+    def create_leann_passages(self, samples: list[dict]):
+        """Create LEANN-compatible passages from LAION data"""
+        print("📝 Creating LEANN passages...")
+
+        passages_file = self.data_dir / "laion_passages.jsonl"
+
+        with open(passages_file, "w", encoding="utf-8") as f:
+            for i, sample in enumerate(samples):
+                passage = {
+                    "id": sample["id"],
+                    "text": sample["caption"],  # Use caption as searchable text
+                    "metadata": {
+                        "image_url": sample["url"],
+                        "image_path": sample.get("image_path", ""),
+                        "width": sample["width"],
+                        "height": sample["height"],
+                        "similarity": sample["similarity"],
+                        "image_index": i,  # Index for embedding lookup
+                    },
+                }
+                f.write(json.dumps(passage) + "\n")
+
+        print(f"✅ Created {len(samples)} passages")
+        return passages_file
+
+    def build_compact_index(
+        self, passages_file: Path, embeddings: np.ndarray, index_path: str, backend: str = "hnsw"
+    ):
+        """Build compact LEANN index with CLIP embeddings (recompute=True, compact=True)"""
+        print(f"🏗️ Building compact LEANN index with {backend} backend...")
+
+        start_time = time.time()
+
+        # Save CLIP embeddings (npy) and also a pickle with (ids, embeddings)
+        npy_path = self.data_dir / "clip_image_embeddings.npy"
+        np.save(npy_path, embeddings)
+        print(f"💾 Saved CLIP embeddings to {npy_path}")
+
+        # Prepare ids in the same order as passages_file (matches embeddings order)
+        ids: list[str] = []
+        with open(passages_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    rec = json.loads(line)
+                    ids.append(str(rec["id"]))
+
+        if len(ids) != embeddings.shape[0]:
+            raise ValueError(
+                f"IDs count ({len(ids)}) does not match embeddings ({embeddings.shape[0]})."
+            )
+
+        pkl_path = self.data_dir / "clip_image_embeddings.pkl"
+        with open(pkl_path, "wb") as pf:
+            pickle.dump((ids, embeddings.astype(np.float32)), pf)
+        print(f"💾 Saved (ids, embeddings) pickle to {pkl_path}")
+
+        # Initialize builder - compact with recompute
+        # Note: For multimodal case, we need to handle embeddings differently
+        # Let's try using sentence-transformers mode but with custom embeddings
+        builder = LeannBuilder(
+            backend_name=backend,
+            # Use CLIP text encoder (ViT-L/14) to match image space (768-dim)
+            embedding_model="clip-ViT-L-14",
+            embedding_mode="sentence-transformers",
+            # HNSW params (or forwarded to chosen backend)
+            graph_degree=32,
+            complexity=64,
+            # Compact/pruned with recompute at query time
+            is_recompute=True,
+            is_compact=True,
+            distance_metric="cosine",  # CLIP uses normalized vectors; cosine is appropriate
+            num_threads=4,
+        )
+
+        # Add passages (text + metadata)
+        print("📚 Adding passages...")
+        self._add_passages_with_embeddings(builder, passages_file, embeddings)
+
+        print(f"🔨 Building compact index at {index_path} from precomputed embeddings...")
+        builder.build_index_from_embeddings(index_path, str(pkl_path))
+
+        build_time = time.time() - start_time
+        print(f"✅ Compact index built in {build_time:.2f}s")
+
+        # Analyze index size
+        self._analyze_index_size(index_path)
+
+        return index_path
+
+    def build_non_compact_index(
+        self, passages_file: Path, embeddings: np.ndarray, index_path: str, backend: str = "hnsw"
+    ):
+        """Build non-compact LEANN index with CLIP embeddings (recompute=False, compact=False)"""
+        print(f"🏗️ Building non-compact LEANN index with {backend} backend...")
+
+        start_time = time.time()
+
+        # Ensure embeddings are saved (npy + pickle)
+        npy_path = self.data_dir / "clip_image_embeddings.npy"
+        if not npy_path.exists():
+            np.save(npy_path, embeddings)
+            print(f"💾 Saved CLIP embeddings to {npy_path}")
+        # Prepare ids in same order as passages_file
+        ids: list[str] = []
+        with open(passages_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    rec = json.loads(line)
+                    ids.append(str(rec["id"]))
+        if len(ids) != embeddings.shape[0]:
+            raise ValueError(
+                f"IDs count ({len(ids)}) does not match embeddings ({embeddings.shape[0]})."
+            )
+        pkl_path = self.data_dir / "clip_image_embeddings.pkl"
+        if not pkl_path.exists():
+            with open(pkl_path, "wb") as pf:
+                pickle.dump((ids, embeddings.astype(np.float32)), pf)
+            print(f"💾 Saved (ids, embeddings) pickle to {pkl_path}")
+
+        # Initialize builder - non-compact without recompute
+        builder = LeannBuilder(
+            backend_name=backend,
+            embedding_model="clip-ViT-L-14",
+            embedding_mode="sentence-transformers",
+            graph_degree=32,
+            complexity=64,
+            is_recompute=False,  # Store embeddings (no recompute needed)
+            is_compact=False,  # Store full index (not pruned)
+            distance_metric="cosine",
+            num_threads=4,
+        )
+
+        # Add passages - embeddings will be loaded from file
+        print("📚 Adding passages...")
+        self._add_passages_with_embeddings(builder, passages_file, embeddings)
+
+        print(f"🔨 Building non-compact index at {index_path} from precomputed embeddings...")
+        builder.build_index_from_embeddings(index_path, str(pkl_path))
+
+        build_time = time.time() - start_time
+        print(f"✅ Non-compact index built in {build_time:.2f}s")
+
+        # Analyze index size
+        self._analyze_index_size(index_path)
+
+        return index_path
+
+    def _add_passages_with_embeddings(self, builder, passages_file: Path, embeddings: np.ndarray):
+        """Helper to add passages with pre-computed CLIP embeddings"""
+        with open(passages_file, encoding="utf-8") as f:
+            for line in tqdm(f, desc="Adding passages"):
+                if line.strip():
+                    passage = json.loads(line)
+
+                    # Add image metadata - LEANN will handle embeddings separately
+                    # Note: We store image metadata and caption text for searchability
+                    # Important: ensure passage ID in metadata matches vector ID
+                    builder.add_text(
+                        text=passage["text"],  # Image caption for searchability
+                        metadata={**passage["metadata"], "id": passage["id"]},
+                    )
+
+    def _analyze_index_size(self, index_path: str):
+        """Analyze index file sizes"""
+        print("📏 Analyzing index sizes...")
+
+        index_path = Path(index_path)
+        index_dir = index_path.parent
+        index_name = index_path.name  # e.g., laion_index.leann
+        index_prefix = index_path.stem  # e.g., laion_index
+
+        files = [
+            (f"{index_prefix}.index", ".index", "core"),
+            (f"{index_name}.meta.json", ".meta.json", "core"),
+            (f"{index_name}.ids.txt", ".ids.txt", "core"),
+            (f"{index_name}.passages.jsonl", ".passages.jsonl", "passages"),
+            (f"{index_name}.passages.idx", ".passages.idx", "passages"),
+        ]
+
+        def _fmt_size(bytes_val: int) -> str:
+            if bytes_val < 1024:
+                return f"{bytes_val} B"
+            kb = bytes_val / 1024
+            if kb < 1024:
+                return f"{kb:.1f} KB"
+            mb = kb / 1024
+            if mb < 1024:
+                return f"{mb:.2f} MB"
+            gb = mb / 1024
+            return f"{gb:.2f} GB"
+
+        total_index_only_mb = 0.0
+        total_all_mb = 0.0
+        for filename, label, group in files:
+            file_path = index_dir / filename
+            if file_path.exists():
+                size_bytes = file_path.stat().st_size
+                print(f"  {label}: {_fmt_size(size_bytes)}")
+                size_mb = size_bytes / (1024 * 1024)
+                total_all_mb += size_mb
+                if group == "core":
+                    total_index_only_mb += size_mb
+            else:
+                print(f"  {label}: (missing)")
+        print(f"  Total (index only, exclude passages): {total_index_only_mb:.2f} MB")
+        print(f"  Total (including passages): {total_all_mb:.2f} MB")
+
+    def create_evaluation_queries(self, samples: list[dict], num_queries: int = 200):
+        """Create evaluation queries from captions"""
+        print(f"📝 Creating {num_queries} evaluation queries...")
+
+        # Sample random captions as queries
+        import random
+
+        random.seed(42)  # For reproducibility
+
+        query_samples = random.sample(samples, min(num_queries, len(samples)))
+
+        queries_file = self.data_dir / "evaluation_queries.jsonl"
+        with open(queries_file, "w", encoding="utf-8") as f:
+            for sample in query_samples:
+                query = {
+                    "id": sample["id"],
+                    "query": sample["caption"],
+                    "ground_truth_id": sample["id"],  # For potential recall evaluation
+                }
+                f.write(json.dumps(query) + "\n")
+
+        print(f"✅ Created {len(query_samples)} evaluation queries")
+        return queries_file
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Setup LAION Multimodal Benchmark")
+    parser.add_argument("--data-dir", default="data", help="Data directory")
+    parser.add_argument("--num-samples", type=int, default=1000, help="Number of LAION samples")
+    parser.add_argument("--num-queries", type=int, default=50, help="Number of evaluation queries")
+    parser.add_argument("--index-path", default="data/laion_index.leann", help="Output index path")
+    parser.add_argument(
+        "--backend", default="hnsw", choices=["hnsw", "diskann"], help="LEANN backend"
+    )
+    parser.add_argument("--skip-download", action="store_true", help="Skip LAION dataset download")
+    parser.add_argument("--skip-build", action="store_true", help="Skip index building")
+
+    args = parser.parse_args()
+
+    print("🚀 Setting up LAION Multimodal Benchmark")
+    print("=" * 50)
+
+    try:
+        # Initialize setup
+        setup = LAIONSetup(args.data_dir)
+
+        # Step 1: Download LAION subset
+        if not args.skip_download:
+            print("\n📦 Step 1: Download LAION subset")
+            samples = setup.download_laion_subset(args.num_samples)
+
+            # Step 2: Generate CLIP image embeddings
+            print("\n🔍 Step 2: Generate CLIP image embeddings")
+            embeddings, valid_samples = setup.generate_clip_image_embeddings(samples)
+
+            # Step 3: Create LEANN passages (image metadata with embeddings)
+            print("\n📝 Step 3: Create LEANN passages")
+            passages_file = setup.create_leann_passages(valid_samples)
+        else:
+            print("⏭️  Skipping LAION dataset download")
+            # Load existing data
+            passages_file = setup.data_dir / "laion_passages.jsonl"
+            embeddings_file = setup.data_dir / "clip_image_embeddings.npy"
+
+            if not passages_file.exists() or not embeddings_file.exists():
+                raise FileNotFoundError(
+                    "Passages or embeddings file not found. Run without --skip-download first."
+                )
+
+            embeddings = np.load(embeddings_file)
+            print(f"📊 Loaded {len(embeddings)} embeddings from {embeddings_file}")
+
+        # Step 4: Build LEANN indexes (both compact and non-compact)
+        if not args.skip_build:
+            print("\n🏗️ Step 4: Build LEANN indexes with CLIP image embeddings")
+
+            # Build compact index (production mode - small, recompute required)
+            compact_index_path = args.index_path
+            print(f"Building compact index: {compact_index_path}")
+            setup.build_compact_index(passages_file, embeddings, compact_index_path, args.backend)
+
+            # Build non-compact index (comparison mode - large, fast search)
+            non_compact_index_path = args.index_path.replace(".leann", "_noncompact.leann")
+            print(f"Building non-compact index: {non_compact_index_path}")
+            setup.build_non_compact_index(
+                passages_file, embeddings, non_compact_index_path, args.backend
+            )
+
+            # Step 5: Build FAISS flat baseline
+            print("\n🔨 Step 5: Build FAISS flat baseline")
+            if not args.skip_download:
+                baseline_path = setup.build_faiss_baseline(embeddings, valid_samples)
+            else:
+                # Load valid_samples from passages file for FAISS baseline
+                valid_samples = []
+                with open(passages_file, encoding="utf-8") as f:
+                    for line in f:
+                        if line.strip():
+                            passage = json.loads(line)
+                            valid_samples.append({"id": passage["id"], "caption": passage["text"]})
+                baseline_path = setup.build_faiss_baseline(embeddings, valid_samples)
+
+            # Step 6: Create evaluation queries
+            print("\n📝 Step 6: Create evaluation queries")
+            queries_file = setup.create_evaluation_queries(valid_samples, args.num_queries)
+        else:
+            print("⏭️  Skipping index building")
+            baseline_path = "data/baseline/faiss_index.bin"
+            queries_file = setup.data_dir / "evaluation_queries.jsonl"
+
+        print("\n🎉 Setup completed successfully!")
+        print("📊 Summary:")
+        if not args.skip_download:
+            print(f"  Downloaded samples: {len(samples)}")
+            print(f"  Valid samples with embeddings: {len(valid_samples)}")
+        else:
+            print(f"  Loaded {len(embeddings)} embeddings")
+
+        if not args.skip_build:
+            print(f"  Compact index: {compact_index_path}")
+            print(f"  Non-compact index: {non_compact_index_path}")
+            print(f"  FAISS baseline: {baseline_path}")
+            print(f"  Queries: {queries_file}")
+
+            print("\n🔧 Next steps:")
+            print(f"  Run evaluation: python evaluate_laion.py --index {compact_index_path}")
+            print(f"  Or compare with: python evaluate_laion.py --index {non_compact_index_path}")
+        else:
+            print("  Skipped building indexes")
+
+    except KeyboardInterrupt:
+        print("\n⚠️  Setup interrupted by user")
+        exit(1)
+    except Exception as e:
+        print(f"\n❌ Setup failed: {e}")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/llm_utils.py

@@ -0,0 +1,342 @@
+"""
+LLM utils for RAG benchmarks with Qwen3-8B and Qwen2.5-VL (multimodal)
+"""
+
+import time
+
+try:
+    import torch
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    HF_AVAILABLE = True
+except ImportError:
+    HF_AVAILABLE = False
+
+try:
+    from vllm import LLM, SamplingParams
+
+    VLLM_AVAILABLE = True
+except ImportError:
+    VLLM_AVAILABLE = False
+
+
+def is_qwen3_model(model_name):
+    """Check if model is Qwen3"""
+    return "Qwen3" in model_name or "qwen3" in model_name.lower()
+
+
+def is_qwen_vl_model(model_name):
+    """Check if model is Qwen2.5-VL"""
+    return "Qwen2.5-VL" in model_name or "qwen2.5-vl" in model_name.lower()
+
+
+def apply_qwen3_chat_template(tokenizer, prompt):
+    """Apply Qwen3 chat template with thinking enabled"""
+    messages = [{"role": "user", "content": prompt}]
+    return tokenizer.apply_chat_template(
+        messages,
+        tokenize=False,
+        add_generation_prompt=True,
+        enable_thinking=True,
+    )
+
+
+def extract_thinking_answer(response):
+    """Extract final answer from Qwen3 thinking model response"""
+    if "<think>" in response and "</think>" in response:
+        try:
+            think_end = response.index("</think>") + len("</think>")
+            final_answer = response[think_end:].strip()
+            return final_answer
+        except (ValueError, IndexError):
+            pass
+
+    return response.strip()
+
+
+def load_hf_model(model_name="Qwen/Qwen3-8B", trust_remote_code=False):
+    """Load HuggingFace model
+
+    Args:
+        model_name (str): Name of the model to load
+        trust_remote_code (bool): Whether to allow execution of code from the model repository.
+            Defaults to False for security. Only enable for trusted models.
+    """
+    if not HF_AVAILABLE:
+        raise ImportError("transformers not available")
+
+    if trust_remote_code:
+        print(
+            "⚠️  WARNING: Loading model with trust_remote_code=True. This can execute arbitrary code."
+        )
+
+    print(f"Loading HF: {model_name}")
+    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=trust_remote_code)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_name,
+        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+        device_map="auto",
+        trust_remote_code=trust_remote_code,
+    )
+    return tokenizer, model
+
+
+def load_vllm_model(model_name="Qwen/Qwen3-8B", trust_remote_code=False):
+    """Load vLLM model
+
+    Args:
+        model_name (str): Name of the model to load
+        trust_remote_code (bool): Whether to allow execution of code from the model repository.
+            Defaults to False for security. Only enable for trusted models.
+    """
+    if not VLLM_AVAILABLE:
+        raise ImportError("vllm not available")
+
+    if trust_remote_code:
+        print(
+            "⚠️  WARNING: Loading model with trust_remote_code=True. This can execute arbitrary code."
+        )
+
+    print(f"Loading vLLM: {model_name}")
+    llm = LLM(model=model_name, trust_remote_code=trust_remote_code)
+
+    # Qwen3 specific config
+    if is_qwen3_model(model_name):
+        stop_tokens = ["<|im_end|>", "<|end_of_text|>"]
+        max_tokens = 2048
+    else:
+        stop_tokens = None
+        max_tokens = 1024
+
+    sampling_params = SamplingParams(temperature=0.7, max_tokens=max_tokens, stop=stop_tokens)
+    return llm, sampling_params
+
+
+def generate_hf(tokenizer, model, prompt, max_tokens=None):
+    """Generate with HF - supports Qwen3 thinking models"""
+    model_name = getattr(model, "name_or_path", "unknown")
+    is_qwen3 = is_qwen3_model(model_name)
+
+    # Apply chat template for Qwen3
+    if is_qwen3:
+        prompt = apply_qwen3_chat_template(tokenizer, prompt)
+        max_tokens = max_tokens or 2048
+    else:
+        max_tokens = max_tokens or 1024
+
+    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=max_tokens,
+            temperature=0.7,
+            do_sample=True,
+            pad_token_id=tokenizer.eos_token_id,
+        )
+    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    response = response[len(prompt) :].strip()
+
+    # Extract final answer for thinking models
+    if is_qwen3:
+        return extract_thinking_answer(response)
+    return response
+
+
+def generate_vllm(llm, sampling_params, prompt):
+    """Generate with vLLM - supports Qwen3 thinking models"""
+    outputs = llm.generate([prompt], sampling_params)
+    response = outputs[0].outputs[0].text.strip()
+
+    # Extract final answer for Qwen3 thinking models
+    model_name = str(llm.llm_engine.model_config.model)
+    if is_qwen3_model(model_name):
+        return extract_thinking_answer(response)
+    return response
+
+
+def create_prompt(context, query, domain="default"):
+    """Create RAG prompt"""
+    if domain == "emails":
+        return f"Email content:\n{context}\n\nQuestion: {query}\n\nAnswer:"
+    elif domain == "finance":
+        return f"Financial content:\n{context}\n\nQuestion: {query}\n\nAnswer:"
+    elif domain == "multimodal":
+        return f"Image context:\n{context}\n\nQuestion: {query}\n\nAnswer:"
+    else:
+        return f"Context: {context}\n\nQuestion: {query}\n\nAnswer:"
+
+
+def evaluate_rag(searcher, llm_func, queries, domain="default", top_k=3, complexity=64):
+    """Simple RAG evaluation with timing"""
+    search_times = []
+    gen_times = []
+    results = []
+
+    for i, query in enumerate(queries):
+        # Search
+        start = time.time()
+        docs = searcher.search(query, top_k=top_k, complexity=complexity)
+        search_time = time.time() - start
+
+        # Generate
+        context = "\n\n".join([doc.text for doc in docs])
+        prompt = create_prompt(context, query, domain)
+
+        start = time.time()
+        response = llm_func(prompt)
+        gen_time = time.time() - start
+
+        search_times.append(search_time)
+        gen_times.append(gen_time)
+        results.append(response)
+
+        if i < 3:
+            print(f"Q{i + 1}: Search={search_time:.3f}s, Gen={gen_time:.3f}s")
+
+    return {
+        "avg_search_time": sum(search_times) / len(search_times),
+        "avg_generation_time": sum(gen_times) / len(gen_times),
+        "results": results,
+    }
+
+
+def load_qwen_vl_model(model_name="Qwen/Qwen2.5-VL-7B-Instruct", trust_remote_code=False):
+    """Load Qwen2.5-VL multimodal model
+
+    Args:
+        model_name (str): Name of the model to load
+        trust_remote_code (bool): Whether to allow execution of code from the model repository.
+            Defaults to False for security. Only enable for trusted models.
+    """
+    if not HF_AVAILABLE:
+        raise ImportError("transformers not available")
+
+    if trust_remote_code:
+        print(
+            "⚠️  WARNING: Loading model with trust_remote_code=True. This can execute arbitrary code."
+        )
+
+    print(f"Loading Qwen2.5-VL: {model_name}")
+
+    try:
+        from transformers import AutoModelForVision2Seq, AutoProcessor
+
+        processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=trust_remote_code)
+        model = AutoModelForVision2Seq.from_pretrained(
+            model_name,
+            torch_dtype=torch.bfloat16,
+            device_map="auto",
+            trust_remote_code=trust_remote_code,
+        )
+
+        return processor, model
+
+    except Exception as e:
+        print(f"Failed to load with AutoModelForVision2Seq, trying specific class: {e}")
+
+        # Fallback to specific class
+        try:
+            from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+
+            processor = AutoProcessor.from_pretrained(
+                model_name, trust_remote_code=trust_remote_code
+            )
+            model = Qwen2VLForConditionalGeneration.from_pretrained(
+                model_name,
+                torch_dtype=torch.bfloat16,
+                device_map="auto",
+                trust_remote_code=trust_remote_code,
+            )
+
+            return processor, model
+
+        except Exception as e2:
+            raise ImportError(f"Failed to load Qwen2.5-VL model: {e2}")
+
+
+def generate_qwen_vl(processor, model, prompt, image_path=None, max_tokens=512):
+    """Generate with Qwen2.5-VL multimodal model"""
+    from PIL import Image
+
+    # Prepare inputs
+    if image_path:
+        image = Image.open(image_path)
+        inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
+    else:
+        inputs = processor(text=prompt, return_tensors="pt").to(model.device)
+
+    # Generate
+    with torch.no_grad():
+        generated_ids = model.generate(
+            **inputs, max_new_tokens=max_tokens, do_sample=False, temperature=0.1
+        )
+
+    # Decode response
+    generated_ids = generated_ids[:, inputs["input_ids"].shape[1] :]
+    response = processor.decode(generated_ids[0], skip_special_tokens=True)
+
+    return response
+
+
+def create_multimodal_prompt(context, query, image_descriptions, task_type="images"):
+    """Create prompt for multimodal RAG"""
+    if task_type == "images":
+        return f"""Based on the retrieved images and their descriptions, answer the following question.
+
+Retrieved Image Descriptions:
+{context}
+
+Question: {query}
+
+Provide a detailed answer based on the visual content described above."""
+
+    return f"Context: {context}\nQuestion: {query}\nAnswer:"
+
+
+def evaluate_multimodal_rag(searcher, queries, processor=None, model=None, complexity=64):
+    """Evaluate multimodal RAG with Qwen2.5-VL"""
+    search_times = []
+    gen_times = []
+    results = []
+
+    for i, query_item in enumerate(queries):
+        # Handle both string and dict formats for queries
+        if isinstance(query_item, dict):
+            query = query_item.get("query", "")
+            image_path = query_item.get("image_path")  # Optional reference image
+        else:
+            query = str(query_item)
+            image_path = None
+
+        # Search
+        start_time = time.time()
+        search_results = searcher.search(query, top_k=3, complexity=complexity)
+        search_time = time.time() - start_time
+        search_times.append(search_time)
+
+        # Prepare context from search results
+        context_parts = []
+        for result in search_results:
+            context_parts.append(f"- {result.text}")
+        context = "\n".join(context_parts)
+
+        # Generate with multimodal model
+        start_time = time.time()
+        if processor and model:
+            prompt = create_multimodal_prompt(context, query, context_parts)
+            response = generate_qwen_vl(processor, model, prompt, image_path)
+        else:
+            response = f"Context: {context}"
+        gen_time = time.time() - start_time
+
+        gen_times.append(gen_time)
+        results.append(response)
+
+        if i < 3:
+            print(f"Q{i + 1}: Search={search_time:.3f}s, Gen={gen_time:.3f}s")
+
+    return {
+        "avg_search_time": sum(search_times) / len(search_times),
+        "avg_generation_time": sum(gen_times) / len(gen_times),
+        "results": results,
+    }

benchmarks/micro_tpt.py

@@ -2,20 +2,20 @@
 
 import argparse
 import time
+from contextlib import contextmanager
 from dataclasses import dataclass
-from typing import Dict, List, Optional, Tuple
 
 import numpy as np
 import torch
 from torch import nn
-from transformers import AutoModel, BitsAndBytesConfig
 from tqdm import tqdm
-from contextlib import contextmanager
+from transformers import AutoModel, BitsAndBytesConfig
+
 
 @dataclass
 class BenchmarkConfig:
     model_path: str
-    batch_sizes: List[int]
+    batch_sizes: list[int]
     seq_length: int
     num_runs: int
     use_fp16: bool = True
@@ -28,47 +28,45 @@ class BenchmarkConfig:
 
 class GraphContainer:
     """Container for managing graphs for different batch sizes (CUDA graphs on NVIDIA, regular on others)."""
-    
+
     def __init__(self, model: nn.Module, seq_length: int):
         self.model = model
         self.seq_length = seq_length
-        self.graphs: Dict[int, 'GraphWrapper'] = {}
-    
-    def get_or_create(self, batch_size: int) -> 'GraphWrapper':
+        self.graphs: dict[int, GraphWrapper] = {}
+
+    def get_or_create(self, batch_size: int) -> "GraphWrapper":
         if batch_size not in self.graphs:
-            self.graphs[batch_size] = GraphWrapper(
-                self.model, batch_size, self.seq_length
-            )
+            self.graphs[batch_size] = GraphWrapper(self.model, batch_size, self.seq_length)
         return self.graphs[batch_size]
 
 
 class GraphWrapper:
     """Wrapper for graph capture and replay (CUDA graphs on NVIDIA, regular on others)."""
-    
+
     def __init__(self, model: nn.Module, batch_size: int, seq_length: int):
         self.model = model
         self.device = self._get_device()
         self.static_input = self._create_random_batch(batch_size, seq_length)
         self.static_attention_mask = torch.ones_like(self.static_input)
-        
+
         # Warm up
         self._warmup()
-        
+
         # Only use CUDA graphs on NVIDIA GPUs
-        if torch.cuda.is_available() and hasattr(torch.cuda, 'CUDAGraph'):
+        if torch.cuda.is_available() and hasattr(torch.cuda, "CUDAGraph"):
             # Capture graph
             self.graph = torch.cuda.CUDAGraph()
             with torch.cuda.graph(self.graph):
                 self.static_output = self.model(
                     input_ids=self.static_input,
-                    attention_mask=self.static_attention_mask
+                    attention_mask=self.static_attention_mask,
                 )
             self.use_cuda_graph = True
         else:
             # For MPS or CPU, just store the model
             self.use_cuda_graph = False
             self.static_output = None
-    
+
     def _get_device(self) -> str:
         if torch.cuda.is_available():
             return "cuda"
@@ -76,22 +74,20 @@ class GraphWrapper:
             return "mps"
         else:
             return "cpu"
-    
+
     def _create_random_batch(self, batch_size: int, seq_length: int) -> torch.Tensor:
         return torch.randint(
-            0, 1000, (batch_size, seq_length), 
-            device=self.device, 
-            dtype=torch.long
+            0, 1000, (batch_size, seq_length), device=self.device, dtype=torch.long
         )
-    
+
     def _warmup(self, num_warmup: int = 3):
         with torch.no_grad():
             for _ in range(num_warmup):
                 self.model(
                     input_ids=self.static_input,
-                    attention_mask=self.static_attention_mask
+                    attention_mask=self.static_attention_mask,
                 )
-    
+
     def __call__(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
         if self.use_cuda_graph:
             self.static_input.copy_(input_ids)
@@ -105,14 +101,14 @@ class GraphWrapper:
 
 class ModelOptimizer:
     """Applies various optimizations to the model."""
-    
+
     @staticmethod
     def optimize(model: nn.Module, config: BenchmarkConfig) -> nn.Module:
         print("\nApplying model optimizations:")
-        
+
         if model is None:
             raise ValueError("Cannot optimize None model")
-        
+
         # Move to GPU
         if torch.cuda.is_available():
             model = model.cuda()
@@ -124,53 +120,59 @@ class ModelOptimizer:
             model = model.cpu()
             device = "cpu"
         print(f"- Model moved to {device}")
-        
+
         # FP16
         if config.use_fp16 and not config.use_int4:
             model = model.half()
             # use torch compile
             model = torch.compile(model)
             print("- Using FP16 precision")
-        
+
         # Check if using SDPA (only on CUDA)
-        if torch.cuda.is_available() and torch.version.cuda and float(torch.version.cuda[:3]) >= 11.6:
-            if hasattr(torch.nn.functional, 'scaled_dot_product_attention'):
+        if (
+            torch.cuda.is_available()
+            and torch.version.cuda
+            and float(torch.version.cuda[:3]) >= 11.6
+        ):
+            if hasattr(torch.nn.functional, "scaled_dot_product_attention"):
                 print("- Using PyTorch SDPA (scaled_dot_product_attention)")
             else:
                 print("- PyTorch SDPA not available")
-        
+
         # Flash Attention (only on CUDA)
         if config.use_flash_attention and torch.cuda.is_available():
             try:
-                from flash_attn.flash_attention import FlashAttention
+                from flash_attn.flash_attention import FlashAttention  # noqa: F401
+
                 print("- Flash Attention 2 available")
                 if hasattr(model.config, "attention_mode"):
                     model.config.attention_mode = "flash_attention_2"
                     print("  - Enabled Flash Attention 2 mode")
             except ImportError:
                 print("- Flash Attention not available")
-        
+
         # Memory efficient attention (only on CUDA)
         if torch.cuda.is_available():
             try:
-                from xformers.ops import memory_efficient_attention
-                if hasattr(model, 'enable_xformers_memory_efficient_attention'):
+                from xformers.ops import memory_efficient_attention  # noqa: F401
+
+                if hasattr(model, "enable_xformers_memory_efficient_attention"):
                     model.enable_xformers_memory_efficient_attention()
                     print("- Enabled xformers memory efficient attention")
                 else:
                     print("- Model doesn't support xformers")
             except (ImportError, AttributeError):
                 print("- Xformers not available")
-        
+
         model.eval()
         print("- Model set to eval mode")
-        
+
         return model
 
 
 class Timer:
     """Handles accurate GPU timing using GPU events or CPU timing."""
-    
+
     def __init__(self):
         if torch.cuda.is_available():
             self.start_event = torch.cuda.Event(enable_timing=True)
@@ -182,7 +184,7 @@ class Timer:
         else:
             # CPU timing
             self.use_gpu_timing = False
-    
+
     @contextmanager
     def timing(self):
         if self.use_gpu_timing:
@@ -195,7 +197,7 @@ class Timer:
             start_time = time.time()
             yield
             self.cpu_elapsed = time.time() - start_time
-    
+
     def elapsed_time(self) -> float:
         if self.use_gpu_timing:
             return self.start_event.elapsed_time(self.end_event) / 1000  # ms to seconds
@@ -205,14 +207,14 @@ class Timer:
 
 class Benchmark:
     """Main benchmark runner."""
-    
+
     def __init__(self, config: BenchmarkConfig):
         self.config = config
         try:
             self.model = self._load_model()
             if self.model is None:
                 raise ValueError("Model initialization failed - model is None")
-            
+
             # Only use CUDA graphs on NVIDIA GPUs
             if config.use_cuda_graphs and torch.cuda.is_available():
                 self.graphs = GraphContainer(self.model, config.seq_length)
@@ -220,25 +222,27 @@ class Benchmark:
                 self.graphs = None
             self.timer = Timer()
         except Exception as e:
-            print(f"ERROR in benchmark initialization: {str(e)}")
+            print(f"ERROR in benchmark initialization: {e!s}")
             raise
-    
+
     def _load_model(self) -> nn.Module:
         print(f"Loading model from {self.config.model_path}...")
-        
+
         try:
             # Int4 quantization using HuggingFace integration
             if self.config.use_int4:
                 import bitsandbytes as bnb
+
                 print(f"- bitsandbytes version: {bnb.__version__}")
-                
-                # 检查是否使用自定义的8bit量化
-                if hasattr(self.config, 'use_linear8bitlt') and self.config.use_linear8bitlt:
+
+                # Check if using custom 8bit quantization
+                if hasattr(self.config, "use_linear8bitlt") and self.config.use_linear8bitlt:
                     print("- Using custom Linear8bitLt replacement for all linear layers")
-                    
-                    # 加载原始模型（不使用量化配置）
+
+                    # Load original model (without quantization config)
                     import bitsandbytes as bnb
                     import torch
+
                     # set default to half
                     torch.set_default_dtype(torch.float16)
                     compute_dtype = torch.float16 if self.config.use_fp16 else torch.float32
@@ -246,112 +250,121 @@ class Benchmark:
                         self.config.model_path,
                         torch_dtype=compute_dtype,
                     )
-                    
-                    # 定义替换函数
+
+                    # Define replacement function
                     def replace_linear_with_linear8bitlt(model):
-                        """递归地将模型中的所有nn.Linear层替换为Linear8bitLt"""
+                        """Recursively replace all nn.Linear layers with Linear8bitLt"""
                         for name, module in list(model.named_children()):
                             if isinstance(module, nn.Linear):
-                                # 获取原始线性层的参数
+                                # Get original linear layer parameters
                                 in_features = module.in_features
                                 out_features = module.out_features
                                 bias = module.bias is not None
-                                
-                                # 创建8bit线性层
+
+                                # Create 8bit linear layer
                                 # print size
                                 print(f"in_features: {in_features}, out_features: {out_features}")
                                 new_module = bnb.nn.Linear8bitLt(
-                                    in_features, 
-                                    out_features, 
-                                    bias=bias, 
-                                    has_fp16_weights=False
+                                    in_features,
+                                    out_features,
+                                    bias=bias,
+                                    has_fp16_weights=False,
                                 )
-                                
-                                # 复制权重和偏置
+
+                                # Copy weights and bias
                                 new_module.weight.data = module.weight.data
                                 if bias:
                                     new_module.bias.data = module.bias.data
-                                    
-                                # 替换模块
+
+                                # Replace module
                                 setattr(model, name, new_module)
                             else:
-                                # 递归处理子模块
+                                # Process child modules recursively
                                 replace_linear_with_linear8bitlt(module)
-                        
+
                         return model
-                    
-                    # 替换所有线性层
+
+                    # Replace all linear layers
                     model = replace_linear_with_linear8bitlt(model)
                     # add torch compile
                     model = torch.compile(model)
-                    
-                    # 将模型移到GPU（量化发生在这里）
-                    device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+
+                    # Move model to GPU (quantization happens here)
+                    device = (
+                        "cuda"
+                        if torch.cuda.is_available()
+                        else "mps"
+                        if torch.backends.mps.is_available()
+                        else "cpu"
+                    )
                     model = model.to(device)
-                    
+
                     print("- All linear layers replaced with Linear8bitLt")
-                    
+
                 else:
-                    # 使用原来的Int4量化方法
+                    # Use original Int4 quantization method
                     print("- Using bitsandbytes for Int4 quantization")
-                    
+
                     # Create quantization config
-                    
+
                     compute_dtype = torch.float16 if self.config.use_fp16 else torch.float32
                     quantization_config = BitsAndBytesConfig(
                         load_in_4bit=True,
                         bnb_4bit_compute_dtype=compute_dtype,
                         bnb_4bit_use_double_quant=True,
-                        bnb_4bit_quant_type="nf4"
+                        bnb_4bit_quant_type="nf4",
                     )
-                    
+
                     print("- Quantization config:", quantization_config)
-                    
+
                     # Load model directly with quantization config
                     model = AutoModel.from_pretrained(
                         self.config.model_path,
                         quantization_config=quantization_config,
                         torch_dtype=compute_dtype,
-                        device_map="auto"  # Let HF decide on device mapping
+                        device_map="auto",  # Let HF decide on device mapping
                     )
-                
+
                 # Check if model loaded successfully
                 if model is None:
                     raise ValueError("Model loading returned None")
-                    
+
                 print(f"- Model type: {type(model)}")
-                
+
                 # Apply optimizations directly here
                 print("\nApplying model optimizations:")
-                
-                if hasattr(self.config, 'use_linear8bitlt') and self.config.use_linear8bitlt:
+
+                if hasattr(self.config, "use_linear8bitlt") and self.config.use_linear8bitlt:
                     print("- Model moved to GPU with Linear8bitLt quantization")
                 else:
                     # Skip moving to GPU since device_map="auto" already did that
                     print("- Model already on GPU due to device_map='auto'")
-                
+
                 # Skip FP16 conversion since we specified compute_dtype
                 print(f"- Using {compute_dtype} for compute dtype")
-                
+
                 # Check CUDA and SDPA
-                if torch.cuda.is_available() and torch.version.cuda and float(torch.version.cuda[:3]) >= 11.6:
-                    if hasattr(torch.nn.functional, 'scaled_dot_product_attention'):
+                if (
+                    torch.cuda.is_available()
+                    and torch.version.cuda
+                    and float(torch.version.cuda[:3]) >= 11.6
+                ):
+                    if hasattr(torch.nn.functional, "scaled_dot_product_attention"):
                         print("- Using PyTorch SDPA (scaled_dot_product_attention)")
                     else:
                         print("- PyTorch SDPA not available")
-                
+
                 # Try xformers if available (only on CUDA)
                 if torch.cuda.is_available():
                     try:
-                        from xformers.ops import memory_efficient_attention
-                        if hasattr(model, 'enable_xformers_memory_efficient_attention'):
+                        if hasattr(model, "enable_xformers_memory_efficient_attention"):
                             model.enable_xformers_memory_efficient_attention()
                             print("- Enabled xformers memory efficient attention")
                         else:
                             print("- Model doesn't support xformers")
                     except (ImportError, AttributeError):
                         print("- Xformers not available")
-                
+
                 # Set to eval mode
                 model.eval()
                 print("- Model set to eval mode")
@@ -365,76 +378,83 @@ class Benchmark:
                     llm_int8_threshold=6.0,
                     llm_int8_has_fp16_weight=False,
                 )
-                
+
                 model = AutoModel.from_pretrained(
                     self.config.model_path,
                     quantization_config=quantization_config,
                     torch_dtype=compute_dtype,
-                    device_map="auto"
+                    device_map="auto",
                 )
-                
+
                 if model is None:
                     raise ValueError("Model loading returned None")
-                    
+
                 print(f"- Model type: {type(model)}")
                 model.eval()
                 print("- Model set to eval mode")
-                
+
             else:
                 # Standard loading for FP16/FP32
                 model = AutoModel.from_pretrained(self.config.model_path)
                 print("- Model loaded in standard precision")
                 print(f"- Model type: {type(model)}")
-                
+
                 # Apply standard optimizations
                 # set default to half
                 import torch
+
                 torch.set_default_dtype(torch.bfloat16)
                 model = ModelOptimizer.optimize(model, self.config)
                 model = model.half()
                 # add torch compile
                 model = torch.compile(model)
-            
+
             # Final check to ensure model is not None
             if model is None:
                 raise ValueError("Model is None after optimization")
-                
+
             print(f"- Final model type: {type(model)}")
             return model
-        
+
         except Exception as e:
-            print(f"ERROR loading model: {str(e)}")
+            print(f"ERROR loading model: {e!s}")
             import traceback
+
             traceback.print_exc()
             raise
-    
+
     def _create_random_batch(self, batch_size: int) -> torch.Tensor:
-        device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+        device = (
+            "cuda"
+            if torch.cuda.is_available()
+            else "mps"
+            if torch.backends.mps.is_available()
+            else "cpu"
+        )
         return torch.randint(
-            0, 1000,
+            0,
+            1000,
             (batch_size, self.config.seq_length),
             device=device,
-            dtype=torch.long
+            dtype=torch.long,
         )
-    
+
     def _run_inference(
-        self,
-        input_ids: torch.Tensor,
-        graph_wrapper: Optional[GraphWrapper] = None
-    ) -> Tuple[float, torch.Tensor]:
+        self, input_ids: torch.Tensor, graph_wrapper: GraphWrapper | None = None
+    ) -> tuple[float, torch.Tensor]:
         attention_mask = torch.ones_like(input_ids)
-        
+
         with torch.no_grad(), self.timer.timing():
             if graph_wrapper is not None:
                 output = graph_wrapper(input_ids, attention_mask)
             else:
                 output = self.model(input_ids=input_ids, attention_mask=attention_mask)
-        
+
         return self.timer.elapsed_time(), output
-    
-    def run(self) -> Dict[int, Dict[str, float]]:
+
+    def run(self) -> dict[int, dict[str, float]]:
         results = {}
-        
+
         # Reset peak memory stats
         if torch.cuda.is_available():
             torch.cuda.reset_peak_memory_stats()
@@ -443,22 +463,20 @@ class Benchmark:
             pass
         else:
             print("- No GPU memory stats available")
-        
+
         for batch_size in self.config.batch_sizes:
             print(f"\nTesting batch size: {batch_size}")
             times = []
-            
+
             # Get or create graph for this batch size
             graph_wrapper = (
-                self.graphs.get_or_create(batch_size)
-                if self.graphs is not None
-                else None
+                self.graphs.get_or_create(batch_size) if self.graphs is not None else None
             )
-            
+
             # Pre-allocate input tensor
             input_ids = self._create_random_batch(batch_size)
             print(f"Input shape: {input_ids.shape}")
-            
+
             # Run benchmark
             for i in tqdm(range(self.config.num_runs), desc=f"Batch size {batch_size}"):
                 try:
@@ -469,44 +487,44 @@ class Benchmark:
                 except Exception as e:
                     print(f"Error during inference: {e}")
                     break
-            
+
             if not times:
                 print(f"No successful runs for batch size {batch_size}, skipping")
                 continue
-                
+
             # Calculate statistics
             avg_time = np.mean(times)
             std_time = np.std(times)
             throughput = batch_size / avg_time
-            
+
             results[batch_size] = {
                 "avg_time": avg_time,
                 "std_time": std_time,
                 "throughput": throughput,
             }
-            
+
             print(f"Avg Time: {avg_time:.4f}s ± {std_time:.4f}s")
             print(f"Throughput: {throughput:.2f} sequences/second")
-        
+
         # Log memory usage
         if torch.cuda.is_available():
-            peak_memory_gb = torch.cuda.max_memory_allocated() / (1024 ** 3)
+            peak_memory_gb = torch.cuda.max_memory_allocated() / (1024**3)
         elif torch.backends.mps.is_available():
             # MPS doesn't have max_memory_allocated, use 0
             peak_memory_gb = 0.0
         else:
             peak_memory_gb = 0.0
             print("- No GPU memory usage available")
-        
+
         if peak_memory_gb > 0:
             print(f"\nPeak GPU memory usage: {peak_memory_gb:.2f} GB")
         else:
             print("\n- GPU memory usage not available")
-        
+
         # Add memory info to results
         for batch_size in results:
             results[batch_size]["peak_memory_gb"] = peak_memory_gb
-        
+
         return results
 
 
@@ -566,14 +584,14 @@ def main():
         action="store_true",
         help="Enable Linear8bitLt quantization for all linear layers",
     )
-    
+
     args = parser.parse_args()
-    
+
     # Print arguments for debugging
     print("\nCommand line arguments:")
     for arg, value in vars(args).items():
         print(f"- {arg}: {value}")
-    
+
     config = BenchmarkConfig(
         model_path=args.model_path,
         batch_sizes=[int(bs) for bs in args.batch_sizes.split(",")],
@@ -586,45 +604,56 @@ def main():
         use_flash_attention=args.use_flash_attention,
         use_linear8bitlt=args.use_linear8bitlt,
     )
-        
+
     # Print configuration for debugging
     print("\nBenchmark configuration:")
     for field, value in vars(config).items():
         print(f"- {field}: {value}")
-    
+
     try:
         benchmark = Benchmark(config)
         results = benchmark.run()
-        
+
         # Save results to file
         import json
         import os
-        
+
         # Create results directory if it doesn't exist
         os.makedirs("results", exist_ok=True)
-        
+
         # Generate filename based on configuration
-        precision_type = "int4" if config.use_int4 else "int8" if config.use_int8 else "fp16" if config.use_fp16 else "fp32"
+        precision_type = (
+            "int4"
+            if config.use_int4
+            else "int8"
+            if config.use_int8
+            else "fp16"
+            if config.use_fp16
+            else "fp32"
+        )
         model_name = os.path.basename(config.model_path)
         output_file = f"results/benchmark_{model_name}_{precision_type}.json"
-        
+
         # Save results
         with open(output_file, "w") as f:
             json.dump(
                 {
-                    "config": {k: str(v) if isinstance(v, list) else v for k, v in vars(config).items()},
-                    "results": {str(k): v for k, v in results.items()}
-                }, 
-                f, 
-                indent=2
+                    "config": {
+                        k: str(v) if isinstance(v, list) else v for k, v in vars(config).items()
+                    },
+                    "results": {str(k): v for k, v in results.items()},
+                },
+                f,
+                indent=2,
             )
         print(f"Results saved to {output_file}")
-        
+
     except Exception as e:
         print(f"Benchmark failed: {e}")
         import traceback
+
         traceback.print_exc()
 
 
 if __name__ == "__main__":
-    main()
+    main()

benchmarks/run_evaluation.py

@@ -5,24 +5,21 @@ It correctly compares results by fetching the text content for both the new sear
 results and the golden standard results, making the comparison robust to ID changes.
 """
 
-import json
 import argparse
+import json
+import sys
 import time
 from pathlib import Path
-import sys
-import numpy as np
-from typing import List
 
-from leann.api import LeannSearcher, LeannBuilder
+import numpy as np
+from leann.api import LeannBuilder, LeannChat, LeannSearcher
 
 
 def download_data_if_needed(data_root: Path, download_embeddings: bool = False):
     """Checks if the data directory exists, and if not, downloads it from HF Hub."""
     if not data_root.exists():
         print(f"Data directory '{data_root}' not found.")
-        print(
-            "Downloading evaluation data from Hugging Face Hub... (this may take a moment)"
-        )
+        print("Downloading evaluation data from Hugging Face Hub... (this may take a moment)")
         try:
             from huggingface_hub import snapshot_download
 
@@ -56,14 +53,14 @@ def download_data_if_needed(data_root: Path, download_embeddings: bool = False):
             print(
                 "Error: huggingface_hub is not installed. Please install it to download the data:"
             )
-            print("uv pip install -e '.[dev]'")
+            print("uv sync --only-group dev")
             sys.exit(1)
         except Exception as e:
             print(f"An error occurred during data download: {e}")
             sys.exit(1)
 
 
-def download_embeddings_if_needed(data_root: Path, dataset_type: str = None):
+def download_embeddings_if_needed(data_root: Path, dataset_type: str | None = None):
     """Download embeddings files specifically."""
     embeddings_dir = data_root / "embeddings"
 
@@ -101,7 +98,7 @@ def download_embeddings_if_needed(data_root: Path, dataset_type: str = None):
 
 
 # --- Helper Function to get Golden Passages ---
-def get_golden_texts(searcher: LeannSearcher, golden_ids: List[int]) -> set:
+def get_golden_texts(searcher: LeannSearcher, golden_ids: list[int]) -> set:
     """
     Retrieves the text for golden passage IDs directly from the LeannSearcher's
     passage manager.
@@ -113,24 +110,20 @@ def get_golden_texts(searcher: LeannSearcher, golden_ids: List[int]) -> set:
             passage_data = searcher.passage_manager.get_passage(str(gid))
             golden_texts.add(passage_data["text"])
         except KeyError:
-            print(
-                f"Warning: Golden passage ID '{gid}' not found in the index's passage data."
-            )
+            print(f"Warning: Golden passage ID '{gid}' not found in the index's passage data.")
     return golden_texts
 
 
-def load_queries(file_path: Path) -> List[str]:
+def load_queries(file_path: Path) -> list[str]:
     queries = []
-    with open(file_path, "r", encoding="utf-8") as f:
+    with open(file_path, encoding="utf-8") as f:
         for line in f:
             data = json.loads(line)
             queries.append(data["query"])
     return queries
 
 
-def build_index_from_embeddings(
-    embeddings_file: str, output_path: str, backend: str = "hnsw"
-):
+def build_index_from_embeddings(embeddings_file: str, output_path: str, backend: str = "hnsw"):
     """
     Build a LEANN index from pre-computed embeddings.
 
@@ -173,9 +166,7 @@ def build_index_from_embeddings(
 
 
 def main():
-    parser = argparse.ArgumentParser(
-        description="Run recall evaluation on a LEANN index."
-    )
+    parser = argparse.ArgumentParser(description="Run recall evaluation on a LEANN index.")
     parser.add_argument(
         "index_path",
         type=str,
@@ -202,26 +193,41 @@ def main():
     parser.add_argument(
         "--num-queries", type=int, default=10, help="Number of queries to evaluate."
     )
-    parser.add_argument(
-        "--top-k", type=int, default=3, help="The 'k' value for recall@k."
-    )
+    parser.add_argument("--top-k", type=int, default=3, help="The 'k' value for recall@k.")
     parser.add_argument(
         "--ef-search", type=int, default=120, help="The 'efSearch' parameter for HNSW."
     )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=0,
+        help="Batch size for HNSW batched search (0 disables batching)",
+    )
+    parser.add_argument(
+        "--llm-type",
+        type=str,
+        choices=["ollama", "hf", "openai", "gemini", "simulated"],
+        default="ollama",
+        help="LLM backend type to optionally query during evaluation (default: ollama)",
+    )
+    parser.add_argument(
+        "--llm-model",
+        type=str,
+        default="qwen3:1.7b",
+        help="LLM model identifier for the chosen backend (default: qwen3:1.7b)",
+    )
     args = parser.parse_args()
 
     # --- Path Configuration ---
-    # Assumes a project structure where the script is in 'examples/'
-    # and data is in 'data/' at the project root.
-    project_root = Path(__file__).resolve().parent.parent
-    data_root = project_root / "data"
+    # Assumes a project structure where the script is in 'benchmarks/'
+    # and evaluation data is in 'benchmarks/data/'.
+    script_dir = Path(__file__).resolve().parent
+    data_root = script_dir / "data"
 
     # Download data based on mode
     if args.mode == "build":
         # For building mode, we need embeddings
-        download_data_if_needed(
-            data_root, download_embeddings=False
-        )  # Basic data first
+        download_data_if_needed(data_root, download_embeddings=False)  # Basic data first
 
         # Auto-detect dataset type and download embeddings
         if args.embeddings_file:
@@ -262,9 +268,7 @@ def main():
         print(f"Index built successfully: {built_index_path}")
 
         # Ask if user wants to run evaluation
-        eval_response = (
-            input("Run evaluation on the built index? (y/n): ").strip().lower()
-        )
+        eval_response = input("Run evaluation on the built index? (y/n): ").strip().lower()
         if eval_response != "y":
             print("Index building complete. Exiting.")
             return
@@ -293,11 +297,9 @@ def main():
                         break
 
             if not args.index_path:
+                print("No indices found. The data download should have included pre-built indices.")
                 print(
-                    "No indices found. The data download should have included pre-built indices."
-                )
-                print(
-                    "Please check the data/indices/ directory or provide --index-path manually."
+                    "Please check the benchmarks/data/indices/ directory or provide --index-path manually."
                 )
                 sys.exit(1)
 
@@ -310,14 +312,10 @@ def main():
     else:
         # Fallback: try to infer from the index directory name
         dataset_type = Path(args.index_path).name
-        print(
-            f"WARNING: Could not detect dataset type from path, inferred '{dataset_type}'."
-        )
+        print(f"WARNING: Could not detect dataset type from path, inferred '{dataset_type}'.")
 
     queries_file = data_root / "queries" / "nq_open.jsonl"
-    golden_results_file = (
-        data_root / "ground_truth" / dataset_type / "flat_results_nq_k3.json"
-    )
+    golden_results_file = data_root / "ground_truth" / dataset_type / "flat_results_nq_k3.json"
 
     print(f"INFO: Detected dataset type: {dataset_type}")
     print(f"INFO: Using queries file: {queries_file}")
@@ -327,7 +325,7 @@ def main():
         searcher = LeannSearcher(args.index_path)
         queries = load_queries(queries_file)
 
-        with open(golden_results_file, "r") as f:
+        with open(golden_results_file) as f:
             golden_results_data = json.load(f)
 
         num_eval_queries = min(args.num_queries, len(queries))
@@ -340,10 +338,23 @@ def main():
         for i in range(num_eval_queries):
             start_time = time.time()
             new_results = searcher.search(
-                queries[i], top_k=args.top_k, ef=args.ef_search
+                queries[i],
+                top_k=args.top_k,
+                complexity=args.ef_search,
+                batch_size=args.batch_size,
             )
             search_times.append(time.time() - start_time)
 
+            # Optional: also call the LLM with configurable backend/model (does not affect recall)
+            llm_config = {"type": args.llm_type, "model": args.llm_model}
+            chat = LeannChat(args.index_path, llm_config=llm_config, searcher=searcher)
+            answer = chat.ask(
+                queries[i],
+                top_k=args.top_k,
+                complexity=args.ef_search,
+                batch_size=args.batch_size,
+            )
+            print(f"Answer: {answer}")
             # Correct Recall Calculation: Based on TEXT content
             new_texts = {result.text for result in new_results}
 

benchmarks/simple_mac_tpt_test.py

@@ -1,26 +1,27 @@
 import time
 from dataclasses import dataclass
-from typing import Dict, List
 
 import numpy as np
 import torch
 from torch import nn
-from transformers import AutoModel, BitsAndBytesConfig
 from tqdm import tqdm
+from transformers import AutoModel
 
 # Add MLX imports
 try:
     import mlx.core as mx
     from mlx_lm.utils import load
+
     MLX_AVAILABLE = True
-except ImportError as e:
+except ImportError:
     print("MLX not available. Install with: uv pip install mlx mlx-lm")
     MLX_AVAILABLE = False
 
+
 @dataclass
 class BenchmarkConfig:
-    model_path: str = "facebook/contriever"
-    batch_sizes: List[int] = None
+    model_path: str = "facebook/contriever-msmarco"
+    batch_sizes: list[int] = None
     seq_length: int = 256
     num_runs: int = 5
     use_fp16: bool = True
@@ -30,18 +31,19 @@ class BenchmarkConfig:
     use_flash_attention: bool = False
     use_linear8bitlt: bool = False
     use_mlx: bool = False  # New flag for MLX testing
-    
+
     def __post_init__(self):
         if self.batch_sizes is None:
-            self.batch_sizes = [1, 2, 4, 8, 16, 32, 64]
+            self.batch_sizes = [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+
 
 class MLXBenchmark:
     """MLX-specific benchmark for embedding models"""
-    
+
     def __init__(self, config: BenchmarkConfig):
         self.config = config
         self.model, self.tokenizer = self._load_model()
-    
+
     def _load_model(self):
         """Load MLX model and tokenizer following the API pattern"""
         print(f"Loading MLX model from {self.config.model_path}...")
@@ -52,55 +54,51 @@ class MLXBenchmark:
         except Exception as e:
             print(f"Error loading MLX model: {e}")
             raise
-    
+
     def _create_random_batch(self, batch_size: int):
         """Create random input batches for MLX testing - same as PyTorch"""
-        return torch.randint(
-            0, 1000,
-            (batch_size, self.config.seq_length),
-            dtype=torch.long
-        )
-    
+        return torch.randint(0, 1000, (batch_size, self.config.seq_length), dtype=torch.long)
+
     def _run_inference(self, input_ids: torch.Tensor) -> float:
         """Run MLX inference with same input as PyTorch"""
         start_time = time.time()
         try:
             # Convert PyTorch tensor to MLX array
             input_ids_mlx = mx.array(input_ids.numpy())
-            
+
             # Get embeddings
             embeddings = self.model(input_ids_mlx)
-            
+
             # Mean pooling (following the API pattern)
             pooled = embeddings.mean(axis=1)
-            
+
             # Convert to numpy (following the API pattern)
             pooled_numpy = np.array(pooled.tolist(), dtype=np.float32)
-            
+
             # Force computation
             _ = pooled_numpy.shape
-            
+
         except Exception as e:
             print(f"MLX inference error: {e}")
-            return float('inf')
+            return float("inf")
         end_time = time.time()
-        
+
         return end_time - start_time
-    
-    def run(self) -> Dict[int, Dict[str, float]]:
+
+    def run(self) -> dict[int, dict[str, float]]:
         """Run the MLX benchmark across all batch sizes"""
         results = {}
-        
+
         print(f"Starting MLX benchmark with model: {self.config.model_path}")
         print(f"Testing batch sizes: {self.config.batch_sizes}")
-        
+
         for batch_size in self.config.batch_sizes:
             print(f"\n=== Testing MLX batch size: {batch_size} ===")
             times = []
-            
+
             # Create input batch (same as PyTorch)
             input_ids = self._create_random_batch(batch_size)
-            
+
             # Warm up
             print("Warming up...")
             for _ in range(3):
@@ -109,26 +107,26 @@ class MLXBenchmark:
                 except Exception as e:
                     print(f"Warmup error: {e}")
                     break
-            
+
             # Run benchmark
-            for i in tqdm(range(self.config.num_runs), desc=f"MLX Batch size {batch_size}"):
+            for _i in tqdm(range(self.config.num_runs), desc=f"MLX Batch size {batch_size}"):
                 try:
                     elapsed_time = self._run_inference(input_ids)
-                    if elapsed_time != float('inf'):
+                    if elapsed_time != float("inf"):
                         times.append(elapsed_time)
                 except Exception as e:
                     print(f"Error during MLX inference: {e}")
                     break
-            
+
             if not times:
                 print(f"Skipping batch size {batch_size} due to errors")
                 continue
-                
+
             # Calculate statistics
             avg_time = np.mean(times)
             std_time = np.std(times)
             throughput = batch_size / avg_time
-            
+
             results[batch_size] = {
                 "avg_time": avg_time,
                 "std_time": std_time,
@@ -136,122 +134,133 @@ class MLXBenchmark:
                 "min_time": np.min(times),
                 "max_time": np.max(times),
             }
-            
+
             print(f"MLX Results for batch size {batch_size}:")
             print(f"  Avg Time: {avg_time:.4f}s ± {std_time:.4f}s")
             print(f"  Min Time: {np.min(times):.4f}s")
             print(f"  Max Time: {np.max(times):.4f}s")
             print(f"  Throughput: {throughput:.2f} sequences/second")
-        
+
         return results
 
+
 class Benchmark:
     def __init__(self, config: BenchmarkConfig):
         self.config = config
-        self.device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+        self.device = (
+            "cuda"
+            if torch.cuda.is_available()
+            else "mps"
+            if torch.backends.mps.is_available()
+            else "cpu"
+        )
         self.model = self._load_model()
-    
+
     def _load_model(self) -> nn.Module:
         print(f"Loading model from {self.config.model_path}...")
-        
-        
+
         model = AutoModel.from_pretrained(self.config.model_path)
         if self.config.use_fp16:
             model = model.half()
         model = torch.compile(model)
         model = model.to(self.device)
-        
+
         model.eval()
         return model
-    
+
     def _create_random_batch(self, batch_size: int) -> torch.Tensor:
         return torch.randint(
-            0, 1000,
+            0,
+            1000,
             (batch_size, self.config.seq_length),
             device=self.device,
-            dtype=torch.long
+            dtype=torch.long,
         )
-    
+
     def _run_inference(self, input_ids: torch.Tensor) -> float:
         attention_mask = torch.ones_like(input_ids)
-        
+        # print shape of input_ids and attention_mask
+        print(f"input_ids shape: {input_ids.shape}")
+        print(f"attention_mask shape: {attention_mask.shape}")
         start_time = time.time()
         with torch.no_grad():
-            output = self.model(input_ids=input_ids, attention_mask=attention_mask)
+            self.model(input_ids=input_ids, attention_mask=attention_mask)
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
+        if torch.backends.mps.is_available():
+            torch.mps.synchronize()
         end_time = time.time()
-        
+
         return end_time - start_time
-    
-    def run(self) -> Dict[int, Dict[str, float]]:
+
+    def run(self) -> dict[int, dict[str, float]]:
         results = {}
-        
+
         if torch.cuda.is_available():
             torch.cuda.reset_peak_memory_stats()
-        
+
         for batch_size in self.config.batch_sizes:
             print(f"\nTesting batch size: {batch_size}")
             times = []
-            
+
             input_ids = self._create_random_batch(batch_size)
-            
-            for i in tqdm(range(self.config.num_runs), desc=f"Batch size {batch_size}"):
+
+            for _i in tqdm(range(self.config.num_runs), desc=f"Batch size {batch_size}"):
                 try:
                     elapsed_time = self._run_inference(input_ids)
                     times.append(elapsed_time)
                 except Exception as e:
                     print(f"Error during inference: {e}")
                     break
-            
+
             if not times:
                 continue
-                
+
             avg_time = np.mean(times)
             std_time = np.std(times)
             throughput = batch_size / avg_time
-            
+
             results[batch_size] = {
                 "avg_time": avg_time,
                 "std_time": std_time,
                 "throughput": throughput,
             }
-            
+
             print(f"Avg Time: {avg_time:.4f}s ± {std_time:.4f}s")
             print(f"Throughput: {throughput:.2f} sequences/second")
-        
+
         if torch.cuda.is_available():
-            peak_memory_gb = torch.cuda.max_memory_allocated() / (1024 ** 3)
+            peak_memory_gb = torch.cuda.max_memory_allocated() / (1024**3)
         else:
             peak_memory_gb = 0.0
-        
+
         for batch_size in results:
             results[batch_size]["peak_memory_gb"] = peak_memory_gb
-        
+
         return results
 
+
 def run_benchmark():
     """Main function to run the benchmark with optimized parameters."""
     config = BenchmarkConfig()
-    
+
     try:
         benchmark = Benchmark(config)
         results = benchmark.run()
-        
+
         max_throughput = max(results[batch_size]["throughput"] for batch_size in results)
         avg_throughput = np.mean([results[batch_size]["throughput"] for batch_size in results])
-        
+
         return {
             "max_throughput": max_throughput,
             "avg_throughput": avg_throughput,
-            "results": results
+            "results": results,
         }
-        
+
     except Exception as e:
         print(f"Benchmark failed: {e}")
-        return {
-            "max_throughput": 0.0,
-            "avg_throughput": 0.0,
-            "error": str(e)
-        }
+        return {"max_throughput": 0.0, "avg_throughput": 0.0, "error": str(e)}
+
 
 def run_mlx_benchmark():
     """Run MLX-specific benchmark"""
@@ -260,55 +269,49 @@ def run_mlx_benchmark():
         return {
             "max_throughput": 0.0,
             "avg_throughput": 0.0,
-            "error": "MLX not available"
+            "error": "MLX not available",
         }
-    
-    config = BenchmarkConfig(
-        model_path="mlx-community/all-MiniLM-L6-v2-4bit",
-        use_mlx=True
-    )
-    
+
+    config = BenchmarkConfig(model_path="mlx-community/all-MiniLM-L6-v2-4bit", use_mlx=True)
+
     try:
         benchmark = MLXBenchmark(config)
         results = benchmark.run()
-        
+
         if not results:
             return {
                 "max_throughput": 0.0,
                 "avg_throughput": 0.0,
-                "error": "No valid results"
+                "error": "No valid results",
             }
-        
+
         max_throughput = max(results[batch_size]["throughput"] for batch_size in results)
         avg_throughput = np.mean([results[batch_size]["throughput"] for batch_size in results])
-        
+
         return {
             "max_throughput": max_throughput,
             "avg_throughput": avg_throughput,
-            "results": results
+            "results": results,
         }
-        
+
     except Exception as e:
         print(f"MLX benchmark failed: {e}")
-        return {
-            "max_throughput": 0.0,
-            "avg_throughput": 0.0,
-            "error": str(e)
-        }
+        return {"max_throughput": 0.0, "avg_throughput": 0.0, "error": str(e)}
+
 
 if __name__ == "__main__":
     print("=== PyTorch Benchmark ===")
     pytorch_result = run_benchmark()
     print(f"PyTorch Max throughput: {pytorch_result['max_throughput']:.2f} sequences/second")
     print(f"PyTorch Average throughput: {pytorch_result['avg_throughput']:.2f} sequences/second")
-    
+
     print("\n=== MLX Benchmark ===")
     mlx_result = run_mlx_benchmark()
     print(f"MLX Max throughput: {mlx_result['max_throughput']:.2f} sequences/second")
     print(f"MLX Average throughput: {mlx_result['avg_throughput']:.2f} sequences/second")
-    
+
     # Compare results
-    if pytorch_result['max_throughput'] > 0 and mlx_result['max_throughput'] > 0:
-        speedup = mlx_result['max_throughput'] / pytorch_result['max_throughput']
-        print(f"\n=== Comparison ===")
-        print(f"MLX is {speedup:.2f}x {'faster' if speedup > 1 else 'slower'} than PyTorch") 
+    if pytorch_result["max_throughput"] > 0 and mlx_result["max_throughput"] > 0:
+        speedup = mlx_result["max_throughput"] / pytorch_result["max_throughput"]
+        print("\n=== Comparison ===")
+        print(f"MLX is {speedup:.2f}x {'faster' if speedup > 1 else 'slower'} than PyTorch")

data/.gitattributes

@@ -1,82 +0,0 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.lz4 filter=lfs diff=lfs merge=lfs -text
-*.mds filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
-# Audio files - uncompressed
-*.pcm filter=lfs diff=lfs merge=lfs -text
-*.sam filter=lfs diff=lfs merge=lfs -text
-*.raw filter=lfs diff=lfs merge=lfs -text
-# Audio files - compressed
-*.aac filter=lfs diff=lfs merge=lfs -text
-*.flac filter=lfs diff=lfs merge=lfs -text
-*.mp3 filter=lfs diff=lfs merge=lfs -text
-*.ogg filter=lfs diff=lfs merge=lfs -text
-*.wav filter=lfs diff=lfs merge=lfs -text
-# Image files - uncompressed
-*.bmp filter=lfs diff=lfs merge=lfs -text
-*.gif filter=lfs diff=lfs merge=lfs -text
-*.png filter=lfs diff=lfs merge=lfs -text
-*.tiff filter=lfs diff=lfs merge=lfs -text
-# Image files - compressed
-*.jpg filter=lfs diff=lfs merge=lfs -text
-*.jpeg filter=lfs diff=lfs merge=lfs -text
-*.webp filter=lfs diff=lfs merge=lfs -text
-# Video files - compressed
-*.mp4 filter=lfs diff=lfs merge=lfs -text
-*.webm filter=lfs diff=lfs merge=lfs -text
-ground_truth/dpr/id_map.json filter=lfs diff=lfs merge=lfs -text
-indices/dpr/dpr_diskann.passages.idx filter=lfs diff=lfs merge=lfs -text
-indices/dpr/dpr_diskann.passages.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/dpr/dpr_diskann_disk.index filter=lfs diff=lfs merge=lfs -text
-indices/dpr/leann.labels.map filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/leann.labels.map filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.index filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.0.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.0.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.1.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.1.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.2.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.2.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.3.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.3.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.4.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.4.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.5.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.5.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.6.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.6.jsonl filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.7.idx filter=lfs diff=lfs merge=lfs -text
-indices/rpj_wiki/rpj_wiki.passages.7.jsonl filter=lfs diff=lfs merge=lfs -text

data/PrideandPrejudice.txt

@@ -1,5 +1,5 @@
 The Project Gutenberg eBook of Pride and Prejudice
-    
+
 This ebook is for the use of anyone anywhere in the United States and
 most other parts of the world at no cost and with almost no restrictions
 whatsoever. You may copy it, give it away or re-use it under the terms
@@ -14557,7 +14557,7 @@ her into Derbyshire, had been the means of uniting them.
 *** END OF THE PROJECT GUTENBERG EBOOK PRIDE AND PREJUDICE ***
 
 
-    
+
 
 Updated editions will replace the previous one—the old editions will
 be renamed.
@@ -14662,7 +14662,7 @@ performed, viewed, copied or distributed:
     at www.gutenberg.org. If you
     are not located in the United States, you will have to check the laws
     of the country where you are located before using this eBook.
-  
+
 1.E.2. If an individual Project Gutenberg™ electronic work is
 derived from texts not protected by U.S. copyright law (does not
 contain a notice indicating that it is posted with permission of the
@@ -14724,7 +14724,7 @@ provided that:
         Gutenberg Literary Archive Foundation at the address specified in
         Section 4, “Information about donations to the Project Gutenberg
         Literary Archive Foundation.”
-    
+
     • You provide a full refund of any money paid by a user who notifies
         you in writing (or by e-mail) within 30 days of receipt that s/he
         does not agree to the terms of the full Project Gutenberg™
@@ -14732,15 +14732,15 @@ provided that:
         copies of the works possessed in a physical medium and discontinue
         all use of and all access to other copies of Project Gutenberg™
         works.
-    
+
     • You provide, in accordance with paragraph 1.F.3, a full refund of
         any money paid for a work or a replacement copy, if a defect in the
         electronic work is discovered and reported to you within 90 days of
         receipt of the work.
-    
+
     • You comply with all other terms of this agreement for free
         distribution of Project Gutenberg™ works.
-    
+
 
 1.E.9. If you wish to charge a fee or distribute a Project
 Gutenberg™ electronic work or group of works on different terms than
@@ -14903,5 +14903,3 @@ This website includes information about Project Gutenberg™,
 including how to make donations to the Project Gutenberg Literary
 Archive Foundation, how to help produce our new eBooks, and how to
 subscribe to our email newsletter to hear about new eBooks.
-
-

demo.ipynb

@@ -4,7 +4,11 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Quick Start in 30s"
+    "# Quick Start \n",
+    "\n",
+    "**Home GitHub Repository:** [LEANN on GitHub](https://github.com/yichuan-w/LEANN)\n",
+    "\n",
+    "**Important for Colab users:** Set your runtime type to T4 GPU for optimal performance. Go to Runtime → Change runtime type → Hardware accelerator → T4 GPU."
    ]
   },
   {
@@ -13,8 +17,25 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# install this if you areusing colab\n",
-    "! pip install leann"
+    "# install this if you are using colab\n",
+    "! uv pip install leann-core leann-backend-hnsw --no-deps\n",
+    "! uv pip install leann --no-deps\n",
+    "# For Colab environment, we need to set some environment variables\n",
+    "import os\n",
+    "\n",
+    "os.environ[\"LEANN_LOG_LEVEL\"] = \"INFO\"  # Enable more detailed logging"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from pathlib import Path\n",
+    "\n",
+    "INDEX_DIR = Path(\"./\").resolve()\n",
+    "INDEX_PATH = str(INDEX_DIR / \"demo.leann\")"
    ]
   },
   {
@@ -26,91 +47,21 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "INFO: Registering backend 'hnsw'\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/Users/yichuan/Desktop/code/LEANN/leann/.venv/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
-      "  from .autonotebook import tqdm as notebook_tqdm\n",
-      "INFO:sentence_transformers.SentenceTransformer:Load pretrained SentenceTransformer: facebook/contriever\n",
-      "WARNING:sentence_transformers.SentenceTransformer:No sentence-transformers model found with name facebook/contriever. Creating a new one with mean pooling.\n",
-      "Writing passages: 100%|██████████| 5/5 [00:00<00:00, 31254.13chunk/s]\n",
-      "Batches: 100%|██████████| 1/1 [00:00<00:00, 12.19it/s]\n",
-      "WARNING:leann_backend_hnsw.hnsw_backend:Converting data to float32, shape: (5, 768)\n",
-      "INFO:leann_backend_hnsw.hnsw_backend:INFO: Converting HNSW index to CSR-pruned format...\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "M: 64 for level: 0\n",
-      "Starting conversion: knowledge.index -> knowledge.csr.tmp\n",
-      "[0.00s] Reading Index HNSW header...\n",
-      "[0.00s]   Header read: d=768, ntotal=5\n",
-      "[0.00s] Reading HNSW struct vectors...\n",
-      "  Reading vector (dtype=<class 'numpy.float64'>, fmt='d')... Count=6, Bytes=48\n",
-      "[0.00s]   Read assign_probas (6)\n",
-      "  Reading vector (dtype=<class 'numpy.int32'>, fmt='i')... Count=7, Bytes=28\n",
-      "[0.11s]   Read cum_nneighbor_per_level (7)\n",
-      "  Reading vector (dtype=<class 'numpy.int32'>, fmt='i')... Count=5, Bytes=20\n",
-      "[0.23s]   Read levels (5)\n",
-      "[0.34s]   Probing for compact storage flag...\n",
-      "[0.34s]   Found compact flag: False\n",
-      "[0.34s]   Compact flag is False, reading original format...\n",
-      "[0.34s]   Probing for potential extra byte before non-compact offsets...\n",
-      "[0.34s]   Found and consumed an unexpected 0x00 byte.\n",
-      "  Reading vector (dtype=<class 'numpy.uint64'>, fmt='Q')... Count=6, Bytes=48\n",
-      "[0.34s]   Read offsets (6)\n",
-      "[0.44s]   Attempting to read neighbors vector...\n",
-      "  Reading vector (dtype=<class 'numpy.int32'>, fmt='i')... Count=320, Bytes=1280\n",
-      "[0.44s]   Read neighbors (320)\n",
-      "[0.54s]   Read scalar params (ep=4, max_lvl=0)\n",
-      "[0.54s] Checking for storage data...\n",
-      "[0.54s]   Found storage fourcc: 49467849.\n",
-      "[0.54s] Converting to CSR format...\n",
-      "[0.54s]   Conversion loop finished.                        \n",
-      "[0.54s] Running validation checks...\n",
-      "    Checking total valid neighbor count...\n",
-      "    OK: Total valid neighbors = 20\n",
-      "    Checking final pointer indices...\n",
-      "    OK: Final pointers match data size.\n",
-      "[0.54s] Deleting original neighbors and offsets arrays...\n",
-      "    CSR Stats: |data|=20, |level_ptr|=10\n",
-      "[0.63s] Writing CSR HNSW graph data in FAISS-compatible order...\n",
-      "   Pruning embeddings: Writing NULL storage marker.\n",
-      "[0.73s] Conversion complete.\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "INFO:leann_backend_hnsw.hnsw_backend:✅ CSR conversion successful.\n",
-      "INFO:leann_backend_hnsw.hnsw_backend:INFO: Replaced original index with CSR-pruned version at 'knowledge.index'\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "from leann.api import LeannBuilder\n",
     "\n",
     "builder = LeannBuilder(backend_name=\"hnsw\")\n",
-    "builder.add_text(\"C# is a powerful programming language\")\n",
-    "builder.add_text(\"Python is a powerful programming language and it is very popular\")\n",
+    "builder.add_text(\"C# is a powerful programming language and it is good at game development\")\n",
+    "builder.add_text(\n",
+    "    \"Python is a powerful programming language and it is good at machine learning tasks\"\n",
+    ")\n",
     "builder.add_text(\"Machine learning transforms industries\")\n",
     "builder.add_text(\"Neural networks process complex data\")\n",
     "builder.add_text(\"Leann is a great storage saving engine for RAG on your MacBook\")\n",
-    "builder.build_index(\"knowledge.leann\")"
+    "builder.build_index(INDEX_PATH)"
    ]
   },
   {
@@ -122,70 +73,13 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "INFO:leann.api:🔍 LeannSearcher.search() called:\n",
-      "INFO:leann.api:  Query: 'programming languages'\n",
-      "INFO:leann.api:  Top_k: 2\n",
-      "INFO:leann.api:  Additional kwargs: {}\n",
-      "INFO:leann.embedding_server_manager:Port 5557 has incompatible server, trying next port...\n",
-      "INFO:leann.embedding_server_manager:Port 5558 has incompatible server, trying next port...\n",
-      "INFO:leann.embedding_server_manager:Port 5559 has incompatible server, trying next port...\n",
-      "INFO:leann.embedding_server_manager:Found compatible server on port 5560\n",
-      "INFO:leann.embedding_server_manager:Using existing compatible server on port 5560\n",
-      "INFO:leann.api:  Launching server time: 0.05758476257324219 seconds\n",
-      "INFO:leann.embedding_server_manager:Found compatible server on port 5560\n",
-      "INFO:leann.embedding_server_manager:Using existing compatible server on port 5560\n",
-      "INFO:leann.api:  Generated embedding shape: (1, 768)\n",
-      "INFO:leann.api:  Embedding time: 0.05983591079711914 seconds\n",
-      "INFO:leann.api:  Search time: 0.039762258529663086 seconds\n",
-      "INFO:leann.api:  Backend returned: labels=2 results\n",
-      "INFO:leann.api:  Processing 2 passage IDs:\n",
-      "INFO:leann.api:    1. passage_id='0' -> SUCCESS: C# is a powerful programming language...\n",
-      "INFO:leann.api:    2. passage_id='1' -> SUCCESS: Python is a powerful programming language and it is very popular...\n",
-      "INFO:leann.api:  Final enriched results: 2 passages\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "[read_HNSW - CSR NL v4] Reading metadata & CSR indices (manual offset)...\n",
-      "[read_HNSW NL v4] Read levels vector, size: 5\n",
-      "[read_HNSW NL v4] Reading Compact Storage format indices...\n",
-      "[read_HNSW NL v4] Read compact_level_ptr, size: 10\n",
-      "[read_HNSW NL v4] Read compact_node_offsets, size: 6\n",
-      "[read_HNSW NL v4] Read entry_point: 4, max_level: 0\n",
-      "[read_HNSW NL v4] Read storage fourcc: 0x6c6c756e\n",
-      "[read_HNSW NL v4 FIX] Detected FileIOReader. Neighbors size field offset: 326\n",
-      "[read_HNSW NL v4] Reading neighbors data into memory.\n",
-      "[read_HNSW NL v4] Read neighbors data, size: 20\n",
-      "[read_HNSW NL v4] Finished reading metadata and CSR indices.\n",
-      "INFO: Skipping external storage loading, since is_recompute is true.\n",
-      "ZmqDistanceComputer initialized: d=768, metric=0\n"
-     ]
-    },
-    {
-     "data": {
-      "text/plain": [
-       "[SearchResult(id='0', score=np.float32(0.9646692), text='C# is a powerful programming language', metadata={}),\n",
-       " SearchResult(id='1', score=np.float32(0.91955304), text='Python is a powerful programming language and it is very popular', metadata={})]"
-      ]
-     },
-     "execution_count": 2,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
+   "outputs": [],
    "source": [
     "from leann.api import LeannSearcher\n",
     "\n",
-    "searcher = LeannSearcher(\"knowledge.leann\")\n",
+    "searcher = LeannSearcher(INDEX_PATH)\n",
     "results = searcher.search(\"programming languages\", top_k=2)\n",
     "results"
    ]
@@ -199,96 +93,22 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "INFO:leann.chat:Attempting to create LLM of type='hf' with model='Qwen/Qwen3-0.6B'\n",
-      "INFO:leann.chat:Initializing HFChat with model='Qwen/Qwen3-0.6B'\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "INFO: Registering backend 'hnsw'\n",
-      "[read_HNSW - CSR NL v4] Reading metadata & CSR indices (manual offset)...\n",
-      "[read_HNSW NL v4] Read levels vector, size: 5\n",
-      "[read_HNSW NL v4] Reading Compact Storage format indices...\n",
-      "[read_HNSW NL v4] Read compact_level_ptr, size: 10\n",
-      "[read_HNSW NL v4] Read compact_node_offsets, size: 6\n",
-      "[read_HNSW NL v4] Read entry_point: 4, max_level: 0\n",
-      "[read_HNSW NL v4] Read storage fourcc: 0x6c6c756e\n",
-      "[read_HNSW NL v4 FIX] Detected FileIOReader. Neighbors size field offset: 326\n",
-      "[read_HNSW NL v4] Reading neighbors data into memory.\n",
-      "[read_HNSW NL v4] Read neighbors data, size: 20\n",
-      "[read_HNSW NL v4] Finished reading metadata and CSR indices.\n",
-      "INFO: Skipping external storage loading, since is_recompute is true.\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/Users/yichuan/Desktop/code/LEANN/leann/.venv/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
-      "  from .autonotebook import tqdm as notebook_tqdm\n",
-      "INFO:leann.chat:MPS is available. Using Apple Silicon GPU.\n",
-      "INFO:leann.api:🔍 LeannSearcher.search() called:\n",
-      "INFO:leann.api:  Query: 'Compare the two retrieved programming languages and say which one is more popular today.'\n",
-      "INFO:leann.api:  Top_k: 2\n",
-      "INFO:leann.api:  Additional kwargs: {}\n",
-      "INFO:leann.embedding_server_manager:Port 5557 has incompatible server, trying next port...\n",
-      "INFO:leann.embedding_server_manager:Port 5558 has incompatible server, trying next port...\n",
-      "INFO:leann.embedding_server_manager:Port 5559 has incompatible server, trying next port...\n",
-      "INFO:leann.embedding_server_manager:Found compatible server on port 5560\n",
-      "INFO:leann.embedding_server_manager:Using existing compatible server on port 5560\n",
-      "INFO:leann.api:  Launching server time: 0.11421084403991699 seconds\n",
-      "INFO:leann.embedding_server_manager:Found compatible server on port 5560\n",
-      "INFO:leann.embedding_server_manager:Using existing compatible server on port 5560\n",
-      "INFO:leann.api:  Generated embedding shape: (1, 768)\n",
-      "INFO:leann.api:  Embedding time: 0.1147918701171875 seconds\n",
-      "INFO:leann.api:  Search time: 0.05468583106994629 seconds\n",
-      "INFO:leann.api:  Backend returned: labels=2 results\n",
-      "INFO:leann.api:  Processing 2 passage IDs:\n",
-      "INFO:leann.api:    1. passage_id='1' -> SUCCESS: Python is a powerful programming language and it is very popular...\n",
-      "INFO:leann.api:    2. passage_id='0' -> SUCCESS: C# is a powerful programming language...\n",
-      "INFO:leann.api:  Final enriched results: 2 passages\n",
-      "INFO:leann.chat:Generating with HuggingFace model, config: {'max_new_tokens': 512, 'temperature': 0.7, 'top_p': 0.9, 'do_sample': True, 'pad_token_id': 151645, 'eos_token_id': 151645}\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "ZmqDistanceComputer initialized: d=768, metric=0\n"
-     ]
-    },
-    {
-     "data": {
-      "text/plain": [
-       "'<think>\\n\\n</think>\\n\\nBased on the context provided, both Python and C# are mentioned as powerful programming languages, but no specific information is given about their popularity today. However, generally, Python is more popular for data science, web development, and other tasks, while C# is widely used in enterprise applications and game development. Since the context does not explicitly state which is more popular, but Python is often considered more popular in many cases, the best answer would be:\\n\\n**Python is more popular today.**'"
-      ]
-     },
-     "execution_count": 1,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
+   "outputs": [],
    "source": [
     "from leann.api import LeannChat\n",
     "\n",
     "llm_config = {\n",
     "    \"type\": \"hf\",\n",
-    "    \"model\": \"Qwen/Qwen3-0.6B\"\n",
+    "    \"model\": \"Qwen/Qwen3-0.6B\",\n",
     "}\n",
     "\n",
-    "chat = LeannChat(index_path=\"knowledge.leann\", llm_config=llm_config)\n",
+    "chat = LeannChat(index_path=INDEX_PATH, llm_config=llm_config)\n",
     "response = chat.ask(\n",
-    "    \"Compare the two retrieved programming languages and say which one is more popular today.\",\n",
+    "    \"Compare the two retrieved programming languages and tell me their advantages.\",\n",
     "    top_k=2,\n",
+    "    llm_kwargs={\"max_tokens\": 128},\n",
     ")\n",
     "response"
    ]

docs/CONTRIBUTING.md

@@ -0,0 +1,223 @@
+# 🤝 Contributing
+
+We welcome contributions! Leann is built by the community, for the community.
+
+## Ways to Contribute
+
+- 🐛 **Bug Reports**: Found an issue? Let us know!
+- 💡 **Feature Requests**: Have an idea? We'd love to hear it!
+- 🔧 **Code Contributions**: PRs welcome for all skill levels
+- 📖 **Documentation**: Help make Leann more accessible
+- 🧪 **Benchmarks**: Share your performance results
+
+## 🚀 Development Setup
+
+### Prerequisites
+
+1. **Install uv** (fast Python package installer):
+   ```bash
+   curl -LsSf https://astral.sh/uv/install.sh | sh
+   ```
+
+2. **Clone the repository**:
+   ```bash
+   git clone https://github.com/LEANN-RAG/LEANN-RAG.git
+   cd LEANN-RAG
+   ```
+
+3. **Install system dependencies**:
+
+   **macOS:**
+   ```bash
+   brew install llvm libomp boost protobuf zeromq pkgconf
+   ```
+
+   **Ubuntu/Debian:**
+   ```bash
+   sudo apt-get install libomp-dev libboost-all-dev protobuf-compiler \
+                        libabsl-dev libmkl-full-dev libaio-dev libzmq3-dev
+   ```
+
+4. **Build from source**:
+   ```bash
+   # macOS
+   CC=$(brew --prefix llvm)/bin/clang CXX=$(brew --prefix llvm)/bin/clang++ uv sync
+
+   # Ubuntu/Debian
+   uv sync
+   ```
+
+## 🔨 Pre-commit Hooks
+
+We use pre-commit hooks to ensure code quality and consistency. This runs automatically before each commit.
+
+### Setup Pre-commit
+
+1. **Install pre-commit tools**:
+   ```bash
+   uv sync lint
+   ```
+
+2. **Install the git hooks**:
+   ```bash
+   pre-commit install
+   ```
+
+3. **Run pre-commit manually** (optional):
+   ```bash
+   uv run pre-commit run --all-files
+   ```
+
+### Pre-commit Checks
+
+Our pre-commit configuration includes:
+- **Trailing whitespace removal**
+- **End-of-file fixing**
+- **YAML validation**
+- **Large file prevention**
+- **Merge conflict detection**
+- **Debug statement detection**
+- **Code formatting with ruff**
+- **Code linting with ruff**
+
+## 🧪 Testing
+
+### Running Tests
+
+```bash
+# Install test tools only (no project runtime)
+uv sync --group test
+
+# Run all tests
+uv run pytest
+
+# Run specific test file
+uv run pytest test/test_filename.py
+
+# Run with coverage
+uv run pytest --cov=leann
+```
+
+### Writing Tests
+
+- Place tests in the `test/` directory
+- Follow the naming convention `test_*.py`
+- Use descriptive test names that explain what's being tested
+- Include both positive and negative test cases
+
+## 📝 Code Style
+
+We use `ruff` for both linting and formatting to ensure consistent code style.
+
+### Format Your Code
+
+```bash
+# Format all files
+ruff format
+
+# Check formatting without changing files
+ruff format --check
+```
+
+### Lint Your Code
+
+```bash
+# Run linter with auto-fix
+ruff check --fix
+
+# Just check without fixing
+ruff check
+```
+
+### Style Guidelines
+
+- Follow PEP 8 conventions
+- Use descriptive variable names
+- Add type hints where appropriate
+- Write docstrings for all public functions and classes
+- Keep functions focused and single-purpose
+
+## 🚦 CI/CD
+
+Our CI pipeline runs automatically on all pull requests. It includes:
+
+1. **Linting and Formatting**: Ensures code follows our style guidelines
+2. **Multi-platform builds**: Tests on Ubuntu and macOS
+3. **Python version matrix**: Tests on Python 3.9-3.13
+4. **Wheel building**: Ensures packages can be built and distributed
+
+### CI Commands
+
+The CI uses the same commands as pre-commit to ensure consistency:
+```bash
+# Linting
+ruff check .
+
+# Format checking
+ruff format --check .
+```
+
+Make sure your code passes these checks locally before pushing!
+
+## 🔄 Pull Request Process
+
+1. **Fork the repository** and create your branch from `main`:
+   ```bash
+   git checkout -b feature/your-feature-name
+   ```
+
+2. **Make your changes**:
+   - Write clean, documented code
+   - Add tests for new functionality
+   - Update documentation as needed
+
+3. **Run pre-commit checks**:
+   ```bash
+   pre-commit run --all-files
+   ```
+
+4. **Test your changes**:
+   ```bash
+   uv run pytest
+   ```
+
+5. **Commit with descriptive messages**:
+   ```bash
+   git commit -m "feat: add new search algorithm"
+   ```
+
+   Follow [Conventional Commits](https://www.conventionalcommits.org/):
+   - `feat:` for new features
+   - `fix:` for bug fixes
+   - `docs:` for documentation changes
+   - `test:` for test additions/changes
+   - `refactor:` for code refactoring
+   - `perf:` for performance improvements
+
+6. **Push and create a pull request**:
+   - Provide a clear description of your changes
+   - Reference any related issues
+   - Include examples or screenshots if applicable
+
+## 📚 Documentation
+
+When adding new features or making significant changes:
+
+1. Update relevant documentation in `/docs`
+2. Add docstrings to new functions/classes
+3. Update README.md if needed
+4. Include usage examples
+
+## 🤔 Getting Help
+
+- **Discord**: Join our community for discussions
+- **Issues**: Check existing issues or create a new one
+- **Discussions**: For general questions and ideas
+
+## 📄 License
+
+By contributing, you agree that your contributions will be licensed under the same license as the project (MIT).
+
+---
+
+Thank you for contributing to LEANN! Every contribution, no matter how small, helps make the project better for everyone. 🌟

docs/RELEASE.md

@@ -1,54 +1,22 @@
 # Release Guide
 
-## One-line Release 🚀
+## Setup (One-time)
 
-```bash
-./scripts/release.sh 0.1.1
-```
+Add `PYPI_API_TOKEN` to GitHub Secrets:
+1. Get token: https://pypi.org/manage/account/token/
+2. Add to secrets: Settings → Secrets → Actions → `PYPI_API_TOKEN`
 
-That's it! This script will:
-1. Update all package versions
-2. Commit and push changes
-3. Create GitHub release
-4. CI automatically builds and publishes to PyPI
+## Release (One-click)
 
-## Manual Testing Before Release
+1. Go to: https://github.com/yichuan-w/LEANN/actions/workflows/release-manual.yml
+2. Click "Run workflow"
+3. Enter version: `0.1.2`
+4. Click green "Run workflow" button
 
-For testing specific packages locally (especially DiskANN on macOS):
+That's it! The workflow will automatically:
+- ✅ Update version in all packages
+- ✅ Build all packages
+- ✅ Publish to PyPI
+- ✅ Create GitHub tag and release
 
-```bash
-# Build specific package locally
-./scripts/build_and_test.sh diskann  # or hnsw, core, meta, all
-
-# Test installation in a clean environment
-python -m venv test_env
-source test_env/bin/activate
-pip install packages/*/dist/*.whl
-
-# Upload to Test PyPI (optional)
-./scripts/upload_to_pypi.sh test
-
-# Upload to Production PyPI (use with caution)
-./scripts/upload_to_pypi.sh prod
-```
-
-### Why Manual Build for DiskANN?
-
-DiskANN's complex dependencies (protobuf, abseil, etc.) sometimes require local testing before release. The build script will:
-- Compile the C++ extension
-- Use `delocate` (macOS) or `auditwheel` (Linux) to bundle system libraries
-- Create a self-contained wheel with no external dependencies
-
-## First-time setup
-
-1. Install GitHub CLI:
-   ```bash
-   brew install gh
-   gh auth login
-   ```
-
-2. Set PyPI token in GitHub:
-   ```bash
-   gh secret set PYPI_API_TOKEN
-   # Paste your PyPI token when prompted
-   ``` 
+Check progress: https://github.com/yichuan-w/LEANN/actions

docs/THINKING_BUDGET_FEATURE.md

@@ -0,0 +1,123 @@
+# Thinking Budget Feature Implementation
+
+## Overview
+
+This document describes the implementation of the **thinking budget** feature for LEANN, which allows users to control the computational effort for reasoning models like GPT-Oss:20b.
+
+## Feature Description
+
+The thinking budget feature provides three levels of computational effort for reasoning models:
+- **`low`**: Fast responses, basic reasoning (default for simple queries)
+- **`medium`**: Balanced speed and reasoning depth
+- **`high`**: Maximum reasoning effort, best for complex analytical questions
+
+## Implementation Details
+
+### 1. Command Line Interface
+
+Added `--thinking-budget` parameter to both CLI and RAG examples:
+
+```bash
+# LEANN CLI
+leann ask my-index --llm ollama --model gpt-oss:20b --thinking-budget high
+
+# RAG Examples
+python apps/email_rag.py --llm ollama --llm-model gpt-oss:20b --thinking-budget high
+python apps/document_rag.py --llm openai --llm-model o3 --thinking-budget medium
+```
+
+### 2. LLM Backend Support
+
+#### Ollama Backend (`packages/leann-core/src/leann/chat.py`)
+
+```python
+def ask(self, prompt: str, **kwargs) -> str:
+    # Handle thinking budget for reasoning models
+    options = kwargs.copy()
+    thinking_budget = kwargs.get("thinking_budget")
+    if thinking_budget:
+        options.pop("thinking_budget", None)
+        if thinking_budget in ["low", "medium", "high"]:
+            options["reasoning"] = {"effort": thinking_budget, "exclude": False}
+```
+
+**API Format**: Uses Ollama's `reasoning` parameter with `effort` and `exclude` fields.
+
+#### OpenAI Backend (`packages/leann-core/src/leann/chat.py`)
+
+```python
+def ask(self, prompt: str, **kwargs) -> str:
+    # Handle thinking budget for reasoning models
+    thinking_budget = kwargs.get("thinking_budget")
+    if thinking_budget and thinking_budget in ["low", "medium", "high"]:
+        # Check if this is an o-series model
+        o_series_models = ["o3", "o3-mini", "o4-mini", "o1", "o3-pro", "o3-deep-research"]
+        if any(model in self.model for model in o_series_models):
+            params["reasoning_effort"] = thinking_budget
+```
+
+**API Format**: Uses OpenAI's `reasoning_effort` parameter for o-series models.
+
+### 3. Parameter Propagation
+
+The thinking budget parameter is properly propagated through the LEANN architecture:
+
+1. **CLI** (`packages/leann-core/src/leann/cli.py`): Captures `--thinking-budget` argument
+2. **Base RAG** (`apps/base_rag_example.py`): Adds parameter to argument parser
+3. **LeannChat** (`packages/leann-core/src/leann/api.py`): Passes `llm_kwargs` to LLM
+4. **LLM Interface**: Handles the parameter in backend-specific implementations
+
+## Files Modified
+
+### Core Implementation
+- `packages/leann-core/src/leann/chat.py`: Added thinking budget support to OllamaChat and OpenAIChat
+- `packages/leann-core/src/leann/cli.py`: Added `--thinking-budget` argument
+- `apps/base_rag_example.py`: Added thinking budget parameter to RAG examples
+
+### Documentation
+- `README.md`: Added thinking budget parameter to usage examples
+- `docs/configuration-guide.md`: Added detailed documentation and usage guidelines
+
+### Examples
+- `examples/thinking_budget_demo.py`: Comprehensive demo script with usage examples
+
+## Usage Examples
+
+### Basic Usage
+```bash
+# High reasoning effort for complex questions
+leann ask my-index --llm ollama --model gpt-oss:20b --thinking-budget high
+
+# Medium reasoning for balanced performance
+leann ask my-index --llm openai --model gpt-4o --thinking-budget medium
+
+# Low reasoning for fast responses
+leann ask my-index --llm ollama --model gpt-oss:20b --thinking-budget low
+```
+
+### RAG Examples
+```bash
+# Email RAG with high reasoning
+python apps/email_rag.py --llm ollama --llm-model gpt-oss:20b --thinking-budget high
+
+# Document RAG with medium reasoning
+python apps/document_rag.py --llm openai --llm-model gpt-4o --thinking-budget medium
+```
+
+## Supported Models
+
+### Ollama Models
+- **GPT-Oss:20b**: Primary target model with reasoning capabilities
+- **Other reasoning models**: Any Ollama model that supports the `reasoning` parameter
+
+### OpenAI Models
+- **o3, o3-mini, o4-mini, o1**: o-series reasoning models with `reasoning_effort` parameter
+- **GPT-OSS models**: Models that support reasoning capabilities
+
+## Testing
+
+The implementation includes comprehensive testing:
+- Parameter handling verification
+- Backend-specific API format validation
+- CLI argument parsing tests
+- Integration with existing LEANN architecture

docs/ast_chunking_guide.md

@@ -0,0 +1,143 @@
+# AST-Aware Code chunking guide
+
+## Overview
+
+This guide covers best practices for using AST-aware code chunking in LEANN. AST chunking provides better semantic understanding of code structure compared to traditional text-based chunking.
+
+## Quick Start
+
+### Basic Usage
+
+```bash
+# Enable AST chunking for mixed content (code + docs)
+python -m apps.document_rag --enable-code-chunking --data-dir ./my_project
+
+# Specialized code repository indexing
+python -m apps.code_rag --repo-dir ./my_codebase
+
+# Global CLI with AST support
+leann build my-code-index --docs ./src --use-ast-chunking
+```
+
+### Installation
+
+```bash
+# Install LEANN with AST chunking support
+uv pip install -e "."
+```
+
+#### For normal users (PyPI install)
+- Use `pip install leann` or `uv pip install leann`.
+- `astchunk` is pulled automatically from PyPI as a dependency; no extra steps.
+
+#### For developers (from source, editable)
+```bash
+git clone https://github.com/yichuan-w/LEANN.git leann
+cd leann
+git submodule update --init --recursive
+uv sync
+```
+- This repo vendors `astchunk` as a git submodule at `packages/astchunk-leann` (our fork).
+- `[tool.uv.sources]` maps the `astchunk` package to that path in editable mode.
+- You can edit code under `packages/astchunk-leann` and Python will use your changes immediately (no separate `pip install astchunk` needed).
+
+## Best Practices
+
+### When to Use AST Chunking
+
+✅ **Recommended for:**
+- Code repositories with multiple languages
+- Mixed documentation and code content
+- Complex codebases with deep function/class hierarchies
+- When working with Claude Code for code assistance
+
+❌ **Not recommended for:**
+- Pure text documents
+- Very large files (>1MB)
+- Languages not supported by tree-sitter
+
+### Optimal Configuration
+
+```bash
+# Recommended settings for most codebases
+python -m apps.code_rag \
+    --repo-dir ./src \
+    --ast-chunk-size 768 \
+    --ast-chunk-overlap 96 \
+    --exclude-dirs .git __pycache__ node_modules build dist
+```
+
+### Supported Languages
+
+| Extension | Language | Status |
+|-----------|----------|--------|
+| `.py` | Python | ✅ Full support |
+| `.java` | Java | ✅ Full support |
+| `.cs` | C# | ✅ Full support |
+| `.ts`, `.tsx` | TypeScript | ✅ Full support |
+| `.js`, `.jsx` | JavaScript | ✅ Via TypeScript parser |
+
+## Integration Examples
+
+### Document RAG with Code Support
+
+```python
+# Enable code chunking in document RAG
+python -m apps.document_rag \
+    --enable-code-chunking \
+    --data-dir ./project \
+    --query "How does authentication work in the codebase?"
+```
+
+### Claude Code Integration
+
+When using with Claude Code MCP server, AST chunking provides better context for:
+- Code completion and suggestions
+- Bug analysis and debugging
+- Architecture understanding
+- Refactoring assistance
+
+## Troubleshooting
+
+### Common Issues
+
+1. **Fallback to Traditional Chunking**
+   - Normal behavior for unsupported languages
+   - Check logs for specific language support
+
+2. **Performance with Large Files**
+   - Adjust `--max-file-size` parameter
+   - Use `--exclude-dirs` to skip unnecessary directories
+
+3. **Quality Issues**
+   - Try different `--ast-chunk-size` values (512, 768, 1024)
+   - Adjust overlap for better context preservation
+
+### Debug Mode
+
+```bash
+export LEANN_LOG_LEVEL=DEBUG
+python -m apps.code_rag --repo-dir ./my_code
+```
+
+## Migration from Traditional Chunking
+
+Existing workflows continue to work without changes. To enable AST chunking:
+
+```bash
+# Before
+python -m apps.document_rag --chunk-size 256
+
+# After (maintains traditional chunking for non-code files)
+python -m apps.document_rag --enable-code-chunking --chunk-size 256 --ast-chunk-size 768
+```
+
+## References
+
+- [astchunk GitHub Repository](https://github.com/yilinjz/astchunk)
+- [LEANN MCP Integration](../packages/leann-mcp/README.md)
+- [Research Paper](https://arxiv.org/html/2506.15655v1)
+
+---
+
+**Note**: AST chunking maintains full backward compatibility while enhancing code understanding capabilities.

docs/code/embedding_model_compare.py

@@ -0,0 +1,98 @@
+"""
+Comparison between Sentence Transformers and OpenAI embeddings
+
+This example shows how different embedding models handle complex queries
+and demonstrates the differences between local and API-based embeddings.
+"""
+
+import numpy as np
+from leann.embedding_compute import compute_embeddings
+
+# OpenAI API key should be set as environment variable
+# export OPENAI_API_KEY="your-api-key-here"
+
+# Test data
+conference_text = "[Title]: COLING 2025 Conference\n[URL]: https://coling2025.org/"
+browser_text = "[Title]: Browser Use Tool\n[URL]: https://github.com/browser-use"
+
+# Two queries with same intent but different wording
+query1 = "Tell me my browser history about some conference i often visit"
+query2 = "browser history about conference I often visit"
+
+texts = [query1, query2, conference_text, browser_text]
+
+
+def cosine_similarity(a, b):
+    return np.dot(a, b)  # Already normalized
+
+
+def analyze_embeddings(embeddings, model_name):
+    print(f"\n=== {model_name} Results ===")
+
+    # Results for Query 1
+    sim1_conf = cosine_similarity(embeddings[0], embeddings[2])
+    sim1_browser = cosine_similarity(embeddings[0], embeddings[3])
+
+    print(f"Query 1: '{query1}'")
+    print(f"  → Conference similarity: {sim1_conf:.4f} {'✓' if sim1_conf > sim1_browser else ''}")
+    print(
+        f"  → Browser similarity:    {sim1_browser:.4f} {'✓' if sim1_browser > sim1_conf else ''}"
+    )
+    print(f"  Winner: {'Conference' if sim1_conf > sim1_browser else 'Browser'}")
+
+    # Results for Query 2
+    sim2_conf = cosine_similarity(embeddings[1], embeddings[2])
+    sim2_browser = cosine_similarity(embeddings[1], embeddings[3])
+
+    print(f"\nQuery 2: '{query2}'")
+    print(f"  → Conference similarity: {sim2_conf:.4f} {'✓' if sim2_conf > sim2_browser else ''}")
+    print(
+        f"  → Browser similarity:    {sim2_browser:.4f} {'✓' if sim2_browser > sim2_conf else ''}"
+    )
+    print(f"  Winner: {'Conference' if sim2_conf > sim2_browser else 'Browser'}")
+
+    # Show the impact
+    print("\n=== Impact Analysis ===")
+    print(f"Conference similarity change: {sim2_conf - sim1_conf:+.4f}")
+    print(f"Browser similarity change:    {sim2_browser - sim1_browser:+.4f}")
+
+    if sim1_conf > sim1_browser and sim2_browser > sim2_conf:
+        print("❌ FLIP: Adding 'browser history' flips winner from Conference to Browser!")
+    elif sim1_conf > sim1_browser and sim2_conf > sim2_browser:
+        print("✅ STABLE: Conference remains winner in both queries")
+    elif sim1_browser > sim1_conf and sim2_browser > sim2_conf:
+        print("✅ STABLE: Browser remains winner in both queries")
+    else:
+        print("🔄 MIXED: Results vary between queries")
+
+    return {
+        "query1_conf": sim1_conf,
+        "query1_browser": sim1_browser,
+        "query2_conf": sim2_conf,
+        "query2_browser": sim2_browser,
+    }
+
+
+# Test Sentence Transformers
+print("Testing Sentence Transformers (facebook/contriever)...")
+try:
+    st_embeddings = compute_embeddings(texts, "facebook/contriever", mode="sentence-transformers")
+    st_results = analyze_embeddings(st_embeddings, "Sentence Transformers (facebook/contriever)")
+except Exception as e:
+    print(f"❌ Sentence Transformers failed: {e}")
+    st_results = None
+
+# Test OpenAI
+print("\n" + "=" * 60)
+print("Testing OpenAI (text-embedding-3-small)...")
+try:
+    openai_embeddings = compute_embeddings(texts, "text-embedding-3-small", mode="openai")
+    openai_results = analyze_embeddings(openai_embeddings, "OpenAI (text-embedding-3-small)")
+except Exception as e:
+    print(f"❌ OpenAI failed: {e}")
+    openai_results = None
+
+# Compare results
+if st_results and openai_results:
+    print("\n" + "=" * 60)
+    print("=== COMPARISON SUMMARY ===")

docs/configuration-guide.md

@@ -0,0 +1,459 @@
+# LEANN Configuration Guide
+
+This guide helps you optimize LEANN for different use cases and understand the trade-offs between various configuration options.
+
+## Getting Started: Simple is Better
+
+When first trying LEANN, start with a small dataset to quickly validate your approach:
+
+**For document RAG**: The default `data/` directory works perfectly - includes 2 AI research papers, Pride and Prejudice literature, and a technical report
+```bash
+python -m apps.document_rag --query "What techniques does LEANN use?"
+```
+
+**For other data sources**: Limit the dataset size for quick testing
+```bash
+# WeChat: Test with recent messages only
+python -m apps.wechat_rag --max-items 100 --query "What did we discuss about the project timeline?"
+
+# Browser history: Last few days
+python -m apps.browser_rag --max-items 500 --query "Find documentation about vector databases"
+
+# Email: Recent inbox
+python -m apps.email_rag --max-items 200 --query "Who sent updates about the deployment status?"
+```
+
+Once validated, scale up gradually:
+- 100 documents → 1,000 → 10,000 → full dataset (`--max-items -1`)
+- This helps identify issues early before committing to long processing times
+
+## Embedding Model Selection: Understanding the Trade-offs
+
+Based on our experience developing LEANN, embedding models fall into three categories:
+
+### Small Models (< 100M parameters)
+**Example**: `sentence-transformers/all-MiniLM-L6-v2` (22M params)
+- **Pros**: Lightweight, fast for both indexing and inference
+- **Cons**: Lower semantic understanding, may miss nuanced relationships
+- **Use when**: Speed is critical, handling simple queries, interactive mode, or just experimenting with LEANN. If time is not a constraint, consider using a larger/better embedding model
+
+### Medium Models (100M-500M parameters)
+**Example**: `facebook/contriever` (110M params), `BAAI/bge-base-en-v1.5` (110M params)
+- **Pros**: Balanced performance, good multilingual support, reasonable speed
+- **Cons**: Requires more compute than small models
+- **Use when**: Need quality results without extreme compute requirements, general-purpose RAG applications
+
+### Large Models (500M+ parameters)
+**Example**: `Qwen/Qwen3-Embedding-0.6B` (600M params), `intfloat/multilingual-e5-large` (560M params)
+- **Pros**: Best semantic understanding, captures complex relationships, excellent multilingual support. **Qwen3-Embedding-0.6B achieves nearly OpenAI API performance!**
+- **Cons**: Slower inference, longer index build times
+- **Use when**: Quality is paramount and you have sufficient compute resources. **Highly recommended** for production use
+
+### Quick Start: Cloud and Local Embedding Options
+
+**OpenAI Embeddings (Fastest Setup)**
+For immediate testing without local model downloads(also if you [do not have GPU](https://github.com/yichuan-w/LEANN/issues/43) and do not care that much about your document leak, you should use this, we compute the embedding and recompute using openai API):
+```bash
+# Set OpenAI embeddings (requires OPENAI_API_KEY)
+--embedding-mode openai --embedding-model text-embedding-3-small
+```
+
+**Ollama Embeddings (Privacy-Focused)**
+For local embeddings with complete privacy:
+```bash
+# First, pull an embedding model
+ollama pull nomic-embed-text
+
+# Use Ollama embeddings
+--embedding-mode ollama --embedding-model nomic-embed-text
+```
+
+<details>
+<summary><strong>Cloud vs Local Trade-offs</strong></summary>
+
+**OpenAI Embeddings** (`text-embedding-3-small/large`)
+- **Pros**: No local compute needed, consistently fast, high quality
+- **Cons**: Requires API key, costs money, data leaves your system, [known limitations with certain languages](https://yichuan-w.github.io/blog/lessons_learned_in_dev_leann/)
+- **When to use**: Prototyping, non-sensitive data, need immediate results
+
+**Local Embeddings**
+- **Pros**: Complete privacy, no ongoing costs, full control, can sometimes outperform OpenAI embeddings
+- **Cons**: Slower than cloud APIs, requires local compute resources
+- **When to use**: Production systems, sensitive data, cost-sensitive applications
+
+</details>
+
+## Local & Remote Inference Endpoints
+
+> Applies to both LLMs (`leann ask`) and embeddings (`leann build`).
+
+LEANN now treats Ollama, LM Studio, and other OpenAI-compatible runtimes as first-class providers. You can point LEANN at any compatible endpoint – either on the same machine or across the network – with a couple of flags or environment variables.
+
+### One-Time Environment Setup
+
+```bash
+# Works for OpenAI-compatible runtimes such as LM Studio, vLLM, SGLang, llamafile, etc.
+export OPENAI_API_KEY="your-key"            # or leave unset for local servers that do not check keys
+export OPENAI_BASE_URL="http://localhost:1234/v1"
+
+# Ollama-compatible runtimes (Ollama, Ollama on another host, llamacpp-server, etc.)
+export LEANN_OLLAMA_HOST="http://localhost:11434"   # falls back to OLLAMA_HOST or LOCAL_LLM_ENDPOINT
+```
+
+LEANN also recognises `LEANN_LOCAL_LLM_HOST` (highest priority), `LEANN_OPENAI_BASE_URL`, and `LOCAL_OPENAI_BASE_URL`, so existing scripts continue to work.
+
+### Passing Hosts Per Command
+
+```bash
+# Build an index with a remote embedding server
+leann build my-notes \
+  --docs ./notes \
+  --embedding-mode openai \
+  --embedding-model text-embedding-qwen3-embedding-0.6b \
+  --embedding-api-base http://192.168.1.50:1234/v1 \
+  --embedding-api-key local-dev-key
+
+# Query using a local LM Studio instance via OpenAI-compatible API
+leann ask my-notes \
+  --llm openai \
+  --llm-model qwen3-8b \
+  --api-base http://localhost:1234/v1 \
+  --api-key local-dev-key
+
+# Query an Ollama instance running on another box
+leann ask my-notes \
+  --llm ollama \
+  --llm-model qwen3:14b \
+  --host http://192.168.1.101:11434
+```
+
+⚠️ **Make sure the endpoint is reachable**: when your inference server runs on a home/workstation and the index/search job runs in the cloud, the server must be able to reach the host you configured. Typical options include:
+
+- Expose a public IP (and open the relevant port) on the machine that hosts LM Studio/Ollama.
+- Configure router or cloud provider port forwarding.
+- Tunnel traffic through tools like `tailscale`, `cloudflared`, or `ssh -R`.
+
+When you set these options while building an index, LEANN stores them in `meta.json`. Any subsequent `leann ask` or searcher process automatically reuses the same provider settings – even when we spawn background embedding servers. This makes the “server without GPU talking to my local workstation” workflow from [issue #80](https://github.com/yichuan-w/LEANN/issues/80#issuecomment-2287230548) work out-of-the-box.
+
+**Tip:** If your runtime does not require an API key (many local stacks don’t), leave `--api-key` unset. LEANN will skip injecting credentials.
+
+### Python API Usage
+
+You can pass the same configuration from Python:
+
+```python
+from leann.api import LeannBuilder
+
+builder = LeannBuilder(
+    backend_name="hnsw",
+    embedding_mode="openai",
+    embedding_model="text-embedding-qwen3-embedding-0.6b",
+    embedding_options={
+        "base_url": "http://192.168.1.50:1234/v1",
+        "api_key": "local-dev-key",
+    },
+)
+builder.build_index("./indexes/my-notes", chunks)
+```
+
+`embedding_options` is persisted to the index `meta.json`, so subsequent `LeannSearcher` or `LeannChat` sessions automatically reuse the same provider settings (the embedding server manager forwards them to the provider for you).
+
+## Index Selection: Matching Your Scale
+
+### HNSW (Hierarchical Navigable Small World)
+**Best for**: Small to medium datasets (< 10M vectors) - **Default and recommended for extreme low storage**
+- Full recomputation required
+- High memory usage during build phase
+- Excellent recall (95%+)
+
+```bash
+# Optimal for most use cases
+--backend-name hnsw --graph-degree 32 --build-complexity 64
+```
+
+### DiskANN
+**Best for**: Large datasets, especially when you want `recompute=True`.
+
+**Key advantages:**
+- **Faster search** on large datasets (3x+ speedup vs HNSW in many cases)
+- **Smart storage**: `recompute=True` enables automatic graph partitioning for smaller indexes
+- **Better scaling**: Designed for 100k+ documents
+
+**Recompute behavior:**
+- `recompute=True` (recommended): Pure PQ traversal + final reranking - faster and enables partitioning
+- `recompute=False`: PQ + partial real distances during traversal - slower but higher accuracy
+
+```bash
+# Recommended for most use cases
+--backend-name diskann --graph-degree 32 --build-complexity 64
+```
+
+**Performance Benchmark**: Run `uv run benchmarks/diskann_vs_hnsw_speed_comparison.py` to compare DiskANN and HNSW on your system.
+
+## LLM Selection: Engine and Model Comparison
+
+### LLM Engines
+
+**OpenAI** (`--llm openai`)
+- **Pros**: Best quality, consistent performance, no local resources needed
+- **Cons**: Costs money ($0.15-2.5 per million tokens), requires internet, data privacy concerns
+- **Models**: `gpt-4o-mini` (fast, cheap), `gpt-4o` (best quality), `o3` (reasoning), `o3-mini` (reasoning, cheaper)
+- **Thinking Budget**: Use `--thinking-budget low/medium/high` for o-series reasoning models (o3, o3-mini, o4-mini)
+- **Note**: Our current default, but we recommend switching to Ollama for most use cases
+
+**Ollama** (`--llm ollama`)
+- **Pros**: Fully local, free, privacy-preserving, good model variety
+- **Cons**: Requires local GPU/CPU resources, slower than cloud APIs, need to install extra [ollama app](https://github.com/ollama/ollama?tab=readme-ov-file#ollama) and pre-download models by `ollama pull`
+- **Models**: `qwen3:0.6b` (ultra-fast), `qwen3:1.7b` (balanced), `qwen3:4b` (good quality), `qwen3:7b` (high quality), `deepseek-r1:1.5b` (reasoning)
+- **Thinking Budget**: Use `--thinking-budget low/medium/high` for reasoning models like GPT-Oss:20b
+
+**HuggingFace** (`--llm hf`)
+- **Pros**: Free tier available, huge model selection, direct model loading (vs Ollama's server-based approach)
+- **Cons**: More complex initial setup
+- **Models**: `Qwen/Qwen3-1.7B-FP8`
+
+## Parameter Tuning Guide
+
+### Search Complexity Parameters
+
+**`--build-complexity`** (index building)
+- Controls thoroughness during index construction
+- Higher = better recall but slower build
+- Recommendations:
+  - 32: Quick prototyping
+  - 64: Balanced (default)
+  - 128: Production systems
+  - 256: Maximum quality
+
+**`--search-complexity`** (query time)
+- Controls search thoroughness
+- Higher = better results but slower
+- Recommendations:
+  - 16: Fast/Interactive search
+  - 32: High quality with diversity
+  - 64+: Maximum accuracy
+
+### Top-K Selection
+
+**`--top-k`** (number of retrieved chunks)
+- More chunks = better context but slower LLM processing
+- Should be always smaller than `--search-complexity`
+- Guidelines:
+  - 10-20: General questions (default: 20)
+  - 30+: Complex multi-hop reasoning requiring comprehensive context
+
+**Trade-off formula**:
+- Retrieval time ∝ log(n) × search_complexity
+- LLM processing time ∝ top_k × chunk_size
+- Total context = top_k × chunk_size tokens
+
+### Thinking Budget for Reasoning Models
+
+**`--thinking-budget`** (reasoning effort level)
+- Controls the computational effort for reasoning models
+- Options: `low`, `medium`, `high`
+- Guidelines:
+  - `low`: Fast responses, basic reasoning (default for simple queries)
+  - `medium`: Balanced speed and reasoning depth
+  - `high`: Maximum reasoning effort, best for complex analytical questions
+- **Supported Models**:
+  - **Ollama**: `gpt-oss:20b`, `gpt-oss:120b`
+  - **OpenAI**: `o3`, `o3-mini`, `o4-mini`, `o1` (o-series reasoning models)
+- **Note**: Models without reasoning support will show a warning and proceed without reasoning parameters
+- **Example**: `--thinking-budget high` for complex analytical questions
+
+**📖 For detailed usage examples and implementation details, check out [Thinking Budget Documentation](THINKING_BUDGET_FEATURE.md)**
+
+**💡 Quick Examples:**
+```bash
+# OpenAI o-series reasoning model
+python apps/document_rag.py --query "What are the main techniques LEANN explores?" \
+  --index-dir hnswbuild --backend hnsw \
+  --llm openai --llm-model o3 --thinking-budget medium
+
+# Ollama reasoning model
+python apps/document_rag.py --query "What are the main techniques LEANN explores?" \
+  --index-dir hnswbuild --backend hnsw \
+  --llm ollama --llm-model gpt-oss:20b --thinking-budget high
+```
+
+### Graph Degree (HNSW/DiskANN)
+
+**`--graph-degree`**
+- Number of connections per node in the graph
+- Higher = better recall but more memory
+- HNSW: 16-32 (default: 32)
+- DiskANN: 32-128 (default: 64)
+
+
+## Performance Optimization Checklist
+
+### If Embedding is Too Slow
+
+1. **Switch to smaller model**:
+   ```bash
+   # From large model
+   --embedding-model Qwen/Qwen3-Embedding-0.6B
+   # To small model
+   --embedding-model sentence-transformers/all-MiniLM-L6-v2
+   ```
+
+2. **Limit dataset size for testing**:
+   ```bash
+   --max-items 1000  # Process first 1k items only
+   ```
+
+3. **Use MLX on Apple Silicon** (optional optimization):
+   ```bash
+   --embedding-mode mlx --embedding-model mlx-community/Qwen3-Embedding-0.6B-8bit
+   ```
+    MLX might not be the best choice, as we tested and found that it only offers 1.3x acceleration compared to HF, so maybe using ollama is a better choice for embedding generation
+
+4. **Use Ollama**
+   ```bash
+   --embedding-mode ollama --embedding-model nomic-embed-text
+   ```
+   To discover additional embedding models in ollama, check out https://ollama.com/search?c=embedding or read more about embedding models at https://ollama.com/blog/embedding-models, please do check the model size that works best for you
+### If Search Quality is Poor
+
+1. **Increase retrieval count**:
+   ```bash
+   --top-k 30  # Retrieve more candidates
+   ```
+
+2. **Upgrade embedding model**:
+   ```bash
+   # For English
+   --embedding-model BAAI/bge-base-en-v1.5
+   # For multilingual
+   --embedding-model intfloat/multilingual-e5-large
+   ```
+
+## Understanding the Trade-offs
+
+Every configuration choice involves trade-offs:
+
+| Factor | Small/Fast | Large/Quality |
+|--------|------------|---------------|
+| Embedding Model | `all-MiniLM-L6-v2` | `Qwen/Qwen3-Embedding-0.6B` |
+| Chunk Size | 512 tokens | 128 tokens |
+| Index Type | HNSW | DiskANN |
+| LLM | `qwen3:1.7b` | `gpt-4o` |
+
+The key is finding the right balance for your specific use case. Start small and simple, measure performance, then scale up only where needed.
+
+## Low-resource setups
+
+If you don’t have a local GPU or builds/searches are too slow, use one or more of the options below.
+
+### 1) Use OpenAI embeddings (no local compute)
+
+Fastest path with zero local GPU requirements. Set your API key and use OpenAI embeddings during build and search:
+
+```bash
+export OPENAI_API_KEY=sk-...
+
+# Build with OpenAI embeddings
+leann build my-index \
+  --embedding-mode openai \
+  --embedding-model text-embedding-3-small
+
+# Search with OpenAI embeddings (recompute at query time)
+leann search my-index "your query" \
+  --recompute
+```
+
+### 2) Run remote builds with SkyPilot (cloud GPU)
+
+Offload embedding generation and index building to a GPU VM using [SkyPilot](https://skypilot.readthedocs.io/en/latest/). A template is provided at `sky/leann-build.yaml`.
+
+```bash
+# One-time: install and configure SkyPilot
+pip install skypilot
+
+# Launch with defaults (L4:1) and mount ./data to ~/leann-data; the build runs automatically
+sky launch -c leann-gpu sky/leann-build.yaml
+
+# Override parameters via -e key=value (optional)
+sky launch -c leann-gpu sky/leann-build.yaml \
+  -e index_name=my-index \
+  -e backend=hnsw \
+  -e embedding_mode=sentence-transformers \
+  -e embedding_model=Qwen/Qwen3-Embedding-0.6B
+
+# Copy the built index back to your local .leann (use rsync)
+rsync -Pavz leann-gpu:~/.leann/indexes/my-index ./.leann/indexes/
+```
+
+### 3) Disable recomputation to trade storage for speed
+
+If you need lower latency and have more storage/memory, disable recomputation. This stores full embeddings and avoids recomputing at search time.
+
+```bash
+# Build without recomputation (HNSW requires non-compact in this mode)
+leann build my-index --no-recompute --no-compact
+
+# Search without recomputation
+leann search my-index "your query" --no-recompute
+```
+
+When to use:
+- Extreme low latency requirements (high QPS, interactive assistants)
+- Read-heavy workloads where storage is cheaper than latency
+- No always-available GPU
+
+Constraints:
+- HNSW: when `--no-recompute` is set, LEANN automatically disables compact mode during build
+- DiskANN: supported; `--no-recompute` skips selective recompute during search
+
+Storage impact:
+- Storing N embeddings of dimension D with float32 requires approximately N × D × 4 bytes
+- Example: 1,000,000 chunks × 768 dims × 4 bytes ≈ 2.86 GB (plus graph/metadata)
+
+Converting an existing index (rebuild required):
+```bash
+# Rebuild in-place (ensure you still have original docs or can regenerate chunks)
+leann build my-index --force --no-recompute --no-compact
+```
+
+Python API usage:
+```python
+from leann import LeannSearcher
+
+searcher = LeannSearcher("/path/to/my-index.leann")
+results = searcher.search("your query", top_k=10, recompute_embeddings=False)
+```
+
+Trade-offs:
+- Lower latency and fewer network hops at query time
+- Significantly higher storage (10–100× vs selective recomputation)
+- Slightly larger memory footprint during build and search
+
+Quick benchmark results (`benchmarks/benchmark_no_recompute.py` with 5k texts, complexity=32):
+
+- HNSW
+
+  ```text
+  recompute=True:  search_time=0.818s, size=1.1MB
+  recompute=False: search_time=0.012s, size=16.6MB
+  ```
+
+- DiskANN
+
+  ```text
+  recompute=True:  search_time=0.041s, size=5.9MB
+  recompute=False: search_time=0.013s, size=24.6MB
+  ```
+
+Conclusion:
+- **HNSW**: `no-recompute` is significantly faster (no embedding recomputation) but requires much more storage (stores all embeddings)
+- **DiskANN**: `no-recompute` uses PQ + partial real distances during traversal (slower but higher accuracy), while `recompute=True` uses pure PQ traversal + final reranking (faster traversal, enables build-time partitioning for smaller storage)
+
+
+
+## Further Reading
+
+- [Lessons Learned Developing LEANN](https://yichuan-w.github.io/blog/lessons_learned_in_dev_leann/)
+- [LEANN Technical Paper](https://arxiv.org/abs/2506.08276)
+- [DiskANN Original Paper](https://suhasjs.github.io/files/diskann_neurips19.pdf)
+- [SSD-based Graph Partitioning](https://github.com/SonglinLife/SSD_BASED_PLAN)

docs/faq.md

@@ -0,0 +1,10 @@
+# FAQ
+
+## 1. My building time seems long
+
+You can speed up the process by using a lightweight embedding model. Add this to your arguments:
+
+```bash
+--embedding-model sentence-transformers/all-MiniLM-L6-v2
+```
+**Model sizes:** `all-MiniLM-L6-v2` (30M parameters), `facebook/contriever` (~100M parameters), `Qwen3-0.6B` (600M parameters)

docs/features.md

@@ -0,0 +1,23 @@
+# ✨ Detailed Features
+
+## 🔥 Core Features
+
+- **🔄 Real-time Embeddings** - Eliminate heavy embedding storage with dynamic computation using optimized ZMQ servers and highly optimized search paradigm (overlapping and batching) with highly optimized embedding engine
+- **🧠 AST-Aware Code Chunking** - Intelligent code chunking that preserves semantic boundaries (functions, classes, methods) for Python, Java, C#, and TypeScript files
+- **📈 Scalable Architecture** - Handles millions of documents on consumer hardware; the larger your dataset, the more LEANN can save
+- **🎯 Graph Pruning** - Advanced techniques to minimize the storage overhead of vector search to a limited footprint
+- **🏗️ Pluggable Backends** - HNSW/FAISS (default), with optional DiskANN for large-scale deployments
+
+## 🛠️ Technical Highlights
+- **🔄 Recompute Mode** - Highest accuracy scenarios while eliminating vector storage overhead
+- **⚡ Zero-copy Operations** - Minimize IPC overhead by transferring distances instead of embeddings
+- **🚀 High-throughput Embedding Pipeline** - Optimized batched processing for maximum efficiency
+- **🎯 Two-level Search** - Novel coarse-to-fine search overlap for accelerated query processing (optional)
+- **💾 Memory-mapped Indices** - Fast startup with raw text mapping to reduce memory overhead
+- **🚀 MLX Support** - Ultra-fast recompute/build with quantized embedding models, accelerating building and search ([minimal example](../examples/mlx_demo.py))
+
+## 🎨 Developer Experience
+
+- **Simple Python API** - Get started in minutes
+- **Extensible backend system** - Easy to add new algorithms
+- **Comprehensive examples** - From basic usage to production deployment

docs/grep_search.md

@@ -0,0 +1,149 @@
+# LEANN Grep Search Usage Guide
+
+## Overview
+
+LEANN's grep search functionality provides exact text matching for finding specific code patterns, error messages, function names, or exact phrases in your indexed documents.
+
+## Basic Usage
+
+### Simple Grep Search
+
+```python
+from leann.api import LeannSearcher
+
+searcher = LeannSearcher("your_index_path")
+
+# Exact text search
+results = searcher.search("def authenticate_user", use_grep=True, top_k=5)
+
+for result in results:
+    print(f"Score: {result.score}")
+    print(f"Text: {result.text[:100]}...")
+    print("-" * 40)
+```
+
+### Comparison: Semantic vs Grep Search
+
+```python
+# Semantic search - finds conceptually similar content
+semantic_results = searcher.search("machine learning algorithms", top_k=3)
+
+# Grep search - finds exact text matches
+grep_results = searcher.search("def train_model", use_grep=True, top_k=3)
+```
+
+## When to Use Grep Search
+
+### Use Cases
+
+- **Code Search**: Finding specific function definitions, class names, or variable references
+- **Error Debugging**: Locating exact error messages or stack traces
+- **Documentation**: Finding specific API endpoints or exact terminology
+
+### Examples
+
+```python
+# Find function definitions
+functions = searcher.search("def __init__", use_grep=True)
+
+# Find import statements
+imports = searcher.search("from sklearn import", use_grep=True)
+
+# Find specific error types
+errors = searcher.search("FileNotFoundError", use_grep=True)
+
+# Find TODO comments
+todos = searcher.search("TODO:", use_grep=True)
+
+# Find configuration entries
+configs = searcher.search("server_port=", use_grep=True)
+```
+
+## Technical Details
+
+### How It Works
+
+1. **File Location**: Grep search operates on the raw text stored in `.jsonl` files
+2. **Command Execution**: Uses the system `grep` command with case-insensitive search
+3. **Result Processing**: Parses JSON lines and extracts text and metadata
+4. **Scoring**: Simple frequency-based scoring based on query term occurrences
+
+### Search Process
+
+```
+Query: "def train_model"
+  ↓
+grep -i -n "def train_model" documents.leann.passages.jsonl
+  ↓
+Parse matching JSON lines
+  ↓
+Calculate scores based on term frequency
+  ↓
+Return top_k results
+```
+
+### Scoring Algorithm
+
+```python
+# Term frequency in document
+score = text.lower().count(query.lower())
+```
+
+Results are ranked by score (highest first), with higher scores indicating more occurrences of the search term.
+
+## Error Handling
+
+### Common Issues
+
+#### Grep Command Not Found
+```
+RuntimeError: grep command not found. Please install grep or use semantic search.
+```
+
+**Solution**: Install grep on your system:
+- **Ubuntu/Debian**: `sudo apt-get install grep`
+- **macOS**: grep is pre-installed
+- **Windows**: Use WSL or install grep via Git Bash/MSYS2
+
+#### No Results Found
+```python
+# Check if your query exists in the raw data
+results = searcher.search("your_query", use_grep=True)
+if not results:
+    print("No exact matches found. Try:")
+    print("1. Check spelling and case")
+    print("2. Use partial terms")
+    print("3. Switch to semantic search")
+```
+
+## Complete Example
+
+```python
+#!/usr/bin/env python3
+"""
+Grep Search Example
+Demonstrates grep search for exact text matching.
+"""
+
+from leann.api import LeannSearcher
+
+def demonstrate_grep_search():
+    # Initialize searcher
+    searcher = LeannSearcher("my_index")
+
+    print("=== Function Search ===")
+    functions = searcher.search("def __init__", use_grep=True, top_k=5)
+    for i, result in enumerate(functions, 1):
+        print(f"{i}. Score: {result.score}")
+        print(f"   Preview: {result.text[:60]}...")
+        print()
+
+    print("=== Error Search ===")
+    errors = searcher.search("FileNotFoundError", use_grep=True, top_k=3)
+    for result in errors:
+        print(f"Content: {result.text.strip()}")
+        print("-" * 40)
+
+if __name__ == "__main__":
+    demonstrate_grep_search()
+```

docs/metadata_filtering.md

@@ -0,0 +1,300 @@
+# LEANN Metadata Filtering Usage Guide
+
+## Overview
+
+Leann possesses metadata filtering capabilities that allow you to filter search results based on arbitrary metadata fields set during chunking. This feature enables use cases like spoiler-free book search, document filtering by date/type, code search by file type, and potentially much more.
+
+## Basic Usage
+
+### Adding Metadata to Your Documents
+
+When building your index, add metadata to each text chunk:
+
+```python
+from leann.api import LeannBuilder
+
+builder = LeannBuilder("hnsw")
+
+# Add text with metadata
+builder.add_text(
+    text="Chapter 1: Alice falls down the rabbit hole",
+    metadata={
+        "chapter": 1,
+        "character": "Alice",
+        "themes": ["adventure", "curiosity"],
+        "word_count": 150
+    }
+)
+
+builder.build_index("alice_in_wonderland_index")
+```
+
+### Searching with Metadata Filters
+
+Use the `metadata_filters` parameter in search calls:
+
+```python
+from leann.api import LeannSearcher
+
+searcher = LeannSearcher("alice_in_wonderland_index")
+
+# Search with filters
+results = searcher.search(
+    query="What happens to Alice?",
+    top_k=10,
+    metadata_filters={
+        "chapter": {"<=": 5},           # Only chapters 1-5
+        "spoiler_level": {"!=": "high"} # No high spoilers
+    }
+)
+```
+
+## Filter Syntax
+
+### Basic Structure
+
+```python
+metadata_filters = {
+    "field_name": {"operator": value},
+    "another_field": {"operator": value}
+}
+```
+
+### Supported Operators
+
+#### Comparison Operators
+- `"=="`: Equal to
+- `"!="`: Not equal to
+- `"<"`: Less than
+- `"<="`: Less than or equal
+- `">"`: Greater than
+- `">="`: Greater than or equal
+
+```python
+# Examples
+{"chapter": {"==": 1}}           # Exactly chapter 1
+{"page": {">": 100}}            # Pages after 100
+{"rating": {">=": 4.0}}         # Rating 4.0 or higher
+{"word_count": {"<": 500}}      # Short passages
+```
+
+#### Membership Operators
+- `"in"`: Value is in list
+- `"not_in"`: Value is not in list
+
+```python
+# Examples
+{"character": {"in": ["Alice", "Bob"]}}      # Alice OR Bob
+{"genre": {"not_in": ["horror", "thriller"]}} # Exclude genres
+{"tags": {"in": ["fiction", "adventure"]}}   # Any of these tags
+```
+
+#### String Operators
+- `"contains"`: String contains substring
+- `"starts_with"`: String starts with prefix
+- `"ends_with"`: String ends with suffix
+
+```python
+# Examples
+{"title": {"contains": "alice"}}        # Title contains "alice"
+{"filename": {"ends_with": ".py"}}      # Python files
+{"author": {"starts_with": "Dr."}}      # Authors with "Dr." prefix
+```
+
+#### Boolean Operators
+- `"is_true"`: Field is truthy
+- `"is_false"`: Field is falsy
+
+```python
+# Examples
+{"is_published": {"is_true": True}}     # Published content
+{"is_draft": {"is_false": False}}       # Not drafts
+```
+
+### Multiple Operators on Same Field
+
+You can apply multiple operators to the same field (AND logic):
+
+```python
+metadata_filters = {
+    "word_count": {
+        ">=": 100,    # At least 100 words
+        "<=": 500     # At most 500 words
+    }
+}
+```
+
+### Compound Filters
+
+Multiple fields are combined with AND logic:
+
+```python
+metadata_filters = {
+    "chapter": {"<=": 10},              # Up to chapter 10
+    "character": {"==": "Alice"},       # About Alice
+    "spoiler_level": {"!=": "high"}     # No major spoilers
+}
+```
+
+## Use Case Examples
+
+### 1. Spoiler-Free Book Search
+
+```python
+# Reader has only read up to chapter 5
+def search_spoiler_free(query, max_chapter):
+    return searcher.search(
+        query=query,
+        metadata_filters={
+            "chapter": {"<=": max_chapter},
+            "spoiler_level": {"in": ["none", "low"]}
+        }
+    )
+
+results = search_spoiler_free("What happens to Alice?", max_chapter=5)
+```
+
+### 2. Document Management by Date
+
+```python
+# Find recent documents
+recent_docs = searcher.search(
+    query="project updates",
+    metadata_filters={
+        "date": {">=": "2024-01-01"},
+        "document_type": {"==": "report"}
+    }
+)
+```
+
+### 3. Code Search by File Type
+
+```python
+# Search only Python files
+python_code = searcher.search(
+    query="authentication function",
+    metadata_filters={
+        "file_extension": {"==": ".py"},
+        "lines_of_code": {"<": 100}
+    }
+)
+```
+
+### 4. Content Filtering by Audience
+
+```python
+# Age-appropriate content
+family_content = searcher.search(
+    query="adventure stories",
+    metadata_filters={
+        "age_rating": {"in": ["G", "PG"]},
+        "content_warnings": {"not_in": ["violence", "adult_themes"]}
+    }
+)
+```
+
+### 5. Multi-Book Series Management
+
+```python
+# Search across first 3 books only
+early_series = searcher.search(
+    query="character development",
+    metadata_filters={
+        "series": {"==": "Harry Potter"},
+        "book_number": {"<=": 3}
+    }
+)
+```
+
+## Running the Example
+
+You can see metadata filtering in action with our spoiler-free book RAG example:
+
+```bash
+# Don't forget to set up the environment
+uv venv
+source .venv/bin/activate
+
+# Set your OpenAI API key (required for embeddings, but you can update the example locally and use ollama instead)
+export OPENAI_API_KEY="your-api-key-here"
+
+# Run the spoiler-free book RAG example
+uv run examples/spoiler_free_book_rag.py
+```
+
+This example demonstrates:
+- Building an index with metadata (chapter numbers, characters, themes, locations)
+- Searching with filters to avoid spoilers (e.g., only show results up to chapter 5)
+- Different scenarios for readers at various points in the book
+
+The example uses Alice's Adventures in Wonderland as sample data and shows how you can search for information without revealing plot points from later chapters.
+
+## Advanced Patterns
+
+### Custom Chunking with metadata
+
+```python
+def chunk_book_with_metadata(book_text, book_info):
+    chunks = []
+
+    for chapter_num, chapter_text in parse_chapters(book_text):
+        # Extract entities, themes, etc.
+        characters = extract_characters(chapter_text)
+        themes = classify_themes(chapter_text)
+        spoiler_level = assess_spoiler_level(chapter_text, chapter_num)
+
+        # Create chunks with rich metadata
+        for paragraph in split_paragraphs(chapter_text):
+            chunks.append({
+                "text": paragraph,
+                "metadata": {
+                    "book_title": book_info["title"],
+                    "chapter": chapter_num,
+                    "characters": characters,
+                    "themes": themes,
+                    "spoiler_level": spoiler_level,
+                    "word_count": len(paragraph.split()),
+                    "reading_level": calculate_reading_level(paragraph)
+                }
+            })
+
+    return chunks
+```
+
+## Performance Considerations
+
+### Efficient Filtering Strategies
+
+1. **Post-search filtering**: Applies filters after vector search, which should be efficient for typical result sets (10-100 results).
+
+2. **Metadata design**: Keep metadata fields simple and avoid deeply nested structures.
+
+### Best Practices
+
+1. **Consistent metadata schema**: Use consistent field names and value types across your documents.
+
+2. **Reasonable metadata size**: Keep metadata reasonably sized to avoid storage overhead.
+
+3. **Type consistency**: Use consistent data types for the same fields (e.g., always integers for chapter numbers).
+
+4. **Index multiple granularities**: Consider chunking at different levels (paragraph, section, chapter) with appropriate metadata.
+
+### Adding Metadata to Existing Indices
+
+To add metadata filtering to existing indices, you'll need to rebuild them with metadata:
+
+```python
+# Read existing passages and add metadata
+def add_metadata_to_existing_chunks(chunks):
+    for chunk in chunks:
+        # Extract or assign metadata based on content
+        chunk["metadata"] = extract_metadata(chunk["text"])
+    return chunks
+
+# Rebuild index with metadata
+enhanced_chunks = add_metadata_to_existing_chunks(existing_chunks)
+builder = LeannBuilder("hnsw")
+for chunk in enhanced_chunks:
+    builder.add_text(chunk["text"], chunk["metadata"])
+builder.build_index("enhanced_index")
+```

docs/normalized_embeddings.md

@@ -0,0 +1,75 @@
+# Normalized Embeddings Support in LEANN
+
+LEANN now automatically detects normalized embedding models and sets the appropriate distance metric for optimal performance.
+
+## What are Normalized Embeddings?
+
+Normalized embeddings are vectors with L2 norm = 1 (unit vectors). These embeddings are optimized for cosine similarity rather than Maximum Inner Product Search (MIPS).
+
+## Automatic Detection
+
+When you create a `LeannBuilder` instance with a normalized embedding model, LEANN will:
+
+1. **Automatically set `distance_metric="cosine"`** if not specified
+2. **Show a warning** if you manually specify a different distance metric
+3. **Provide optimal search performance** with the correct metric
+
+## Supported Normalized Embedding Models
+
+### OpenAI
+All OpenAI text embedding models are normalized:
+- `text-embedding-ada-002`
+- `text-embedding-3-small`
+- `text-embedding-3-large`
+
+### Voyage AI
+All Voyage AI embedding models are normalized:
+- `voyage-2`
+- `voyage-3`
+- `voyage-large-2`
+- `voyage-multilingual-2`
+- `voyage-code-2`
+
+### Cohere
+All Cohere embedding models are normalized:
+- `embed-english-v3.0`
+- `embed-multilingual-v3.0`
+- `embed-english-light-v3.0`
+- `embed-multilingual-light-v3.0`
+
+## Example Usage
+
+```python
+from leann.api import LeannBuilder
+
+# Automatic detection - will use cosine distance
+builder = LeannBuilder(
+    backend_name="hnsw",
+    embedding_model="text-embedding-3-small",
+    embedding_mode="openai"
+)
+# Warning: Detected normalized embeddings model 'text-embedding-3-small'...
+# Automatically setting distance_metric='cosine'
+
+# Manual override (not recommended)
+builder = LeannBuilder(
+    backend_name="hnsw",
+    embedding_model="text-embedding-3-small",
+    embedding_mode="openai",
+    distance_metric="mips"  # Will show warning
+)
+# Warning: Using 'mips' distance metric with normalized embeddings...
+```
+
+## Non-Normalized Embeddings
+
+Models like `facebook/contriever` and other sentence-transformers models that are not normalized will continue to use MIPS by default, which is optimal for them.
+
+## Why This Matters
+
+Using the wrong distance metric with normalized embeddings can lead to:
+- **Poor search quality** due to HNSW's early termination with narrow score ranges
+- **Incorrect ranking** of search results
+- **Suboptimal performance** compared to using the correct metric
+
+For more details on why this happens, see our analysis in the [embedding detection code](../packages/leann-core/src/leann/api.py) which automatically handles normalized embeddings and MIPS distance metric issues.

docs/roadmap.md

@@ -0,0 +1,21 @@
+# 📈 Roadmap
+
+## 🎯 Q2 2025
+
+- [X] HNSW backend integration
+- [X] DiskANN backend with MIPS/L2/Cosine support
+- [X] Real-time embedding pipeline
+- [X] Memory-efficient graph pruning
+
+## 🚀 Q3 2025
+
+- [ ] Advanced caching strategies
+- [ ] Add contextual-retrieval https://www.anthropic.com/news/contextual-retrieval
+- [ ] Add sleep-time-compute and summarize agent! to summarilze the file on computer!
+- [ ] Add OpenAI recompute API
+
+## 🌟 Q4 2025
+
+- [ ] Integration with LangChain/LlamaIndex
+- [ ] Visual similarity search
+- [ ] Query rewrtiting, rerank and expansion

examples/basic_demo.py

@@ -1,21 +1,28 @@
 """
 Simple demo showing basic leann usage
-Run: uv run python examples/simple_demo.py
+Run: uv run python examples/basic_demo.py
 """
 
 import argparse
-from leann import LeannBuilder, LeannSearcher, LeannChat
+
+from leann import LeannBuilder, LeannChat, LeannSearcher
 
 
 def main():
-    parser = argparse.ArgumentParser(description="Simple demo of Leann with selectable embedding models.")
-    parser.add_argument("--embedding_model", type=str, default="sentence-transformers/all-mpnet-base-v2",
-                        help="The embedding model to use, e.g., 'sentence-transformers/all-mpnet-base-v2' or 'text-embedding-ada-002'.")
+    parser = argparse.ArgumentParser(
+        description="Simple demo of Leann with selectable embedding models."
+    )
+    parser.add_argument(
+        "--embedding_model",
+        type=str,
+        default="sentence-transformers/all-mpnet-base-v2",
+        help="The embedding model to use, e.g., 'sentence-transformers/all-mpnet-base-v2' or 'text-embedding-ada-002'.",
+    )
     args = parser.parse_args()
 
     print(f"=== Leann Simple Demo with {args.embedding_model} ===")
     print()
-    
+
     # Sample knowledge base
     chunks = [
         "Machine learning is a subset of artificial intelligence that enables computers to learn without being explicitly programmed.",
@@ -27,7 +34,7 @@ def main():
         "Big data refers to extremely large datasets that require special tools and techniques to process.",
         "Cloud computing provides on-demand access to computing resources over the internet.",
     ]
-    
+
     print("1. Building index (no embeddings stored)...")
     builder = LeannBuilder(
         embedding_model=args.embedding_model,
@@ -37,45 +44,45 @@ def main():
         builder.add_text(chunk)
     builder.build_index("demo_knowledge.leann")
     print()
-    
+
     print("2. Searching with real-time embeddings...")
     searcher = LeannSearcher("demo_knowledge.leann")
-    
+
     queries = [
         "What is machine learning?",
-        "How does neural network work?", 
+        "How does neural network work?",
         "Tell me about data processing",
     ]
-    
+
     for query in queries:
         print(f"Query: {query}")
         results = searcher.search(query, top_k=2)
-        
+
         for i, result in enumerate(results, 1):
             print(f"  {i}. Score: {result.score:.3f}")
             print(f"     Text: {result.text[:100]}...")
         print()
-    
+
     print("3. Interactive chat demo:")
     print("   (Note: Requires OpenAI API key for real responses)")
-    
+
     chat = LeannChat("demo_knowledge.leann")
-    
+
     # Demo questions
     demo_questions: list[str] = [
         "What is the difference between machine learning and deep learning?",
         "How is data science related to big data?",
     ]
-    
+
     for question in demo_questions:
         print(f"   Q: {question}")
         response = chat.ask(question)
         print(f"   A: {response}")
         print()
-    
+
     print("Demo completed! Try running:")
-    print("   uv run python examples/document_search.py")
+    print("   uv run python apps/document_rag.py")
 
 
 if __name__ == "__main__":
-    main()
+    main()

examples/document_search.py

@@ -1,146 +0,0 @@
-#!/usr/bin/env python3
-"""
-Document search demo with recompute mode
-"""
-
-import os
-from pathlib import Path
-import shutil
-import time
-
-# Import backend packages to trigger plugin registration
-try:
-    import leann_backend_diskann
-    import leann_backend_hnsw
-    print("INFO: Backend packages imported successfully.")
-except ImportError as e:
-    print(f"WARNING: Could not import backend packages. Error: {e}")
-
-# Import upper-level API from leann-core
-from leann.api import LeannBuilder, LeannSearcher, LeannChat
-
-
-def load_sample_documents():
-    """Create sample documents for demonstration"""
-    docs = [
-        {"title": "Intro to Python", "content": "Python is a high-level, interpreted language known for simplicity."},
-        {"title": "ML Basics", "content": "Machine learning builds systems that learn from data."},
-        {"title": "Data Structures", "content": "Data structures like arrays, lists, and graphs organize data."},
-    ]
-    return docs
-
-def main():
-    print("==========================================================")
-    print("=== Leann Document Search Demo (DiskANN + Recompute) ===")
-    print("==========================================================")
-    
-    INDEX_DIR = Path("./test_indices")
-    INDEX_PATH = str(INDEX_DIR / "documents.diskann")
-    BACKEND_TO_TEST = "diskann"
-
-    if INDEX_DIR.exists():
-        print(f"--- Cleaning up old index directory: {INDEX_DIR} ---")
-        shutil.rmtree(INDEX_DIR)
-
-    # --- 1. Build index ---
-    print(f"\n[PHASE 1] Building index using '{BACKEND_TO_TEST}' backend...")
-    
-    builder = LeannBuilder(
-        backend_name=BACKEND_TO_TEST, 
-        graph_degree=32, 
-        complexity=64
-    )
-    
-    documents = load_sample_documents()
-    print(f"Loaded {len(documents)} sample documents.")
-    for doc in documents:
-        builder.add_text(doc["content"], metadata={"title": doc["title"]})
-        
-    builder.build_index(INDEX_PATH)
-    print(f"\nIndex built!")
-
-    # --- 2. Basic search demo ---
-    print(f"\n[PHASE 2] Basic search using '{BACKEND_TO_TEST}' backend...")
-    searcher = LeannSearcher(index_path=INDEX_PATH)
-    
-    query = "What is machine learning?"
-    print(f"\nQuery: '{query}'")
-    
-    print("\n--- Basic search mode (PQ computation) ---")
-    start_time = time.time()
-    results = searcher.search(query, top_k=2)
-    basic_time = time.time() - start_time
-    
-    print(f"⏱️  Basic search time: {basic_time:.3f} seconds")
-    print(">>> Basic search results <<<")
-    for i, res in enumerate(results, 1):
-        print(f"  {i}. ID: {res.id}, Score: {res.score:.4f}, Text: '{res.text}', Metadata: {res.metadata}")
-
-    # --- 3. Recompute search demo ---
-    print(f"\n[PHASE 3] Recompute search using embedding server...")
-    
-    print("\n--- Recompute search mode (get real embeddings via network) ---")
-    
-    # Configure recompute parameters
-    recompute_params = {
-        "recompute_beighbor_embeddings": True,  # Enable network recomputation
-        "USE_DEFERRED_FETCH": False,           # Don't use deferred fetch
-        "skip_search_reorder": True,           # Skip search reordering
-        "dedup_node_dis": True,               # Enable node distance deduplication
-        "prune_ratio": 0.1,                   # Pruning ratio 10%
-        "batch_recompute": False,             # Don't use batch recomputation
-        "global_pruning": False,              # Don't use global pruning
-        "zmq_port": 5555,                     # ZMQ port
-        "embedding_model": "sentence-transformers/all-mpnet-base-v2"
-    }
-    
-    print("Recompute parameter configuration:")
-    for key, value in recompute_params.items():
-        print(f"  {key}: {value}")
-    
-    print(f"\n🔄 Executing Recompute search...")
-    try:
-        start_time = time.time()
-        recompute_results = searcher.search(query, top_k=2, **recompute_params)
-        recompute_time = time.time() - start_time
-        
-        print(f"⏱️  Recompute search time: {recompute_time:.3f} seconds")
-        print(">>> Recompute search results <<<")
-        for i, res in enumerate(recompute_results, 1):
-            print(f"  {i}. ID: {res.id}, Score: {res.score:.4f}, Text: '{res.text}', Metadata: {res.metadata}")
-        
-        # Compare results
-        print(f"\n--- Result comparison ---")
-        print(f"Basic search time: {basic_time:.3f} seconds")
-        print(f"Recompute time: {recompute_time:.3f} seconds")
-        
-        print("\nBasic search vs Recompute results:")
-        for i in range(min(len(results), len(recompute_results))):
-            basic_score = results[i].score
-            recompute_score = recompute_results[i].score
-            score_diff = abs(basic_score - recompute_score)
-            print(f"  Position {i+1}: PQ={basic_score:.4f}, Recompute={recompute_score:.4f}, Difference={score_diff:.4f}")
-        
-        if recompute_time > basic_time:
-            print(f"✅ Recompute mode working correctly (more accurate but slower)")
-        else:
-            print(f"ℹ️  Recompute time is unusually fast, network recomputation may not be enabled")
-            
-    except Exception as e:
-        print(f"❌ Recompute search failed: {e}")
-        print("This usually indicates an embedding server connection issue")
-
-    # --- 4. Chat demo ---
-    print(f"\n[PHASE 4] Starting chat session...")
-    chat = LeannChat(index_path=INDEX_PATH)
-    chat_response = chat.ask(query)
-    print(f"You: {query}")
-    print(f"Leann: {chat_response}")
-
-    print("\n==========================================================")
-    print("✅ Demo finished successfully!")
-    print("==========================================================")
-
-
-if __name__ == "__main__":
-    main()

examples/dynamic_update_no_recompute.py

@@ -0,0 +1,429 @@
+"""Dynamic HNSW update demo without compact storage.
+
+This script reproduces the minimal scenario we used while debugging on-the-fly
+recompute:
+
+1. Build a non-compact HNSW index from the first few paragraphs of a text file.
+2. Print the top results with `recompute_embeddings=True`.
+3. Append additional paragraphs with :meth:`LeannBuilder.update_index`.
+4. Run the same query again to show the newly inserted passages.
+
+Run it with ``uv`` (optionally pointing LEANN_HNSW_LOG_PATH at a file to inspect
+ZMQ activity)::
+
+    LEANN_HNSW_LOG_PATH=embedding_fetch.log \
+    uv run -m examples.dynamic_update_no_recompute \
+      --index-path .leann/examples/leann-demo.leann
+
+By default the script builds an index from ``data/2501.14312v1 (1).pdf`` and
+then updates it with LEANN-related material from ``data/2506.08276v1.pdf``.
+It issues the query "What's LEANN?" before and after the update to show how the
+new passages become immediately searchable. The script uses the
+``sentence-transformers/all-MiniLM-L6-v2`` model with ``is_recompute=True`` so
+Faiss pulls existing vectors on demand via the ZMQ embedding server, while
+freshly added passages are embedded locally just like the initial build.
+
+To make storage comparisons easy, the script can also build a matching
+``is_recompute=False`` baseline (enabled by default) and report the index size
+delta after the update. Disable the baseline run with
+``--skip-compare-no-recompute`` if you only need the recompute flow.
+"""
+
+import argparse
+import json
+from collections.abc import Iterable
+from pathlib import Path
+from typing import Any
+
+from leann.api import LeannBuilder, LeannSearcher
+from leann.registry import register_project_directory
+
+from apps.chunking import create_text_chunks
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+
+DEFAULT_QUERY = "What's LEANN?"
+DEFAULT_INITIAL_FILES = [
+    REPO_ROOT / "data" / "2501.14312v1 (1).pdf",
+    REPO_ROOT / "data" / "huawei_pangu.md",
+    REPO_ROOT / "data" / "PrideandPrejudice.txt",
+]
+DEFAULT_UPDATE_FILES = [REPO_ROOT / "data" / "2506.08276v1.pdf"]
+
+
+def load_chunks_from_files(paths: list[Path]) -> list[str]:
+    from llama_index.core import SimpleDirectoryReader
+
+    documents = []
+    for path in paths:
+        p = path.expanduser().resolve()
+        if not p.exists():
+            raise FileNotFoundError(f"Input path not found: {p}")
+        if p.is_dir():
+            reader = SimpleDirectoryReader(str(p), recursive=False)
+            documents.extend(reader.load_data(show_progress=True))
+        else:
+            reader = SimpleDirectoryReader(input_files=[str(p)])
+            documents.extend(reader.load_data(show_progress=True))
+
+    if not documents:
+        return []
+
+    chunks = create_text_chunks(
+        documents,
+        chunk_size=512,
+        chunk_overlap=128,
+        use_ast_chunking=False,
+    )
+    return [c for c in chunks if isinstance(c, str) and c.strip()]
+
+
+def run_search(index_path: Path, query: str, top_k: int, *, recompute_embeddings: bool) -> list:
+    searcher = LeannSearcher(str(index_path))
+    try:
+        return searcher.search(
+            query=query,
+            top_k=top_k,
+            recompute_embeddings=recompute_embeddings,
+            batch_size=16,
+        )
+    finally:
+        searcher.cleanup()
+
+
+def print_results(title: str, results: Iterable) -> None:
+    print(f"\n=== {title} ===")
+    res_list = list(results)
+    print(f"results count: {len(res_list)}")
+    print("passages:")
+    if not res_list:
+        print("  (no passages returned)")
+    for res in res_list:
+        snippet = res.text.replace("\n", " ")[:120]
+        print(f"  - {res.id}: {snippet}... (score={res.score:.4f})")
+
+
+def build_initial_index(
+    index_path: Path,
+    paragraphs: list[str],
+    model_name: str,
+    embedding_mode: str,
+    is_recompute: bool,
+) -> None:
+    builder = LeannBuilder(
+        backend_name="hnsw",
+        embedding_model=model_name,
+        embedding_mode=embedding_mode,
+        is_compact=False,
+        is_recompute=is_recompute,
+    )
+    for idx, passage in enumerate(paragraphs):
+        builder.add_text(passage, metadata={"id": str(idx)})
+    builder.build_index(str(index_path))
+
+
+def update_index(
+    index_path: Path,
+    start_id: int,
+    paragraphs: list[str],
+    model_name: str,
+    embedding_mode: str,
+    is_recompute: bool,
+) -> None:
+    updater = LeannBuilder(
+        backend_name="hnsw",
+        embedding_model=model_name,
+        embedding_mode=embedding_mode,
+        is_compact=False,
+        is_recompute=is_recompute,
+    )
+    for offset, passage in enumerate(paragraphs, start=start_id):
+        updater.add_text(passage, metadata={"id": str(offset)})
+    updater.update_index(str(index_path))
+
+
+def ensure_index_dir(index_path: Path) -> None:
+    index_path.parent.mkdir(parents=True, exist_ok=True)
+
+
+def cleanup_index_files(index_path: Path) -> None:
+    """Remove leftover index artifacts for a clean rebuild."""
+
+    parent = index_path.parent
+    if not parent.exists():
+        return
+    stem = index_path.stem
+    for file in parent.glob(f"{stem}*"):
+        if file.is_file():
+            file.unlink()
+
+
+def index_file_size(index_path: Path) -> int:
+    """Return the size of the primary .index file for the given index path."""
+
+    index_file = index_path.parent / f"{index_path.stem}.index"
+    return index_file.stat().st_size if index_file.exists() else 0
+
+
+def load_metadata_snapshot(index_path: Path) -> dict[str, Any] | None:
+    meta_path = index_path.parent / f"{index_path.name}.meta.json"
+    if not meta_path.exists():
+        return None
+    try:
+        return json.loads(meta_path.read_text())
+    except json.JSONDecodeError:
+        return None
+
+
+def run_workflow(
+    *,
+    label: str,
+    index_path: Path,
+    initial_paragraphs: list[str],
+    update_paragraphs: list[str],
+    model_name: str,
+    embedding_mode: str,
+    is_recompute: bool,
+    query: str,
+    top_k: int,
+    skip_search: bool,
+) -> dict[str, Any]:
+    prefix = f"[{label}] " if label else ""
+
+    ensure_index_dir(index_path)
+    cleanup_index_files(index_path)
+
+    print(f"{prefix}Building initial index...")
+    build_initial_index(
+        index_path,
+        initial_paragraphs,
+        model_name,
+        embedding_mode,
+        is_recompute=is_recompute,
+    )
+
+    initial_size = index_file_size(index_path)
+    if not skip_search:
+        before_results = run_search(
+            index_path,
+            query,
+            top_k,
+            recompute_embeddings=is_recompute,
+        )
+    else:
+        before_results = None
+
+    print(f"\n{prefix}Updating index with additional passages...")
+    update_index(
+        index_path,
+        start_id=len(initial_paragraphs),
+        paragraphs=update_paragraphs,
+        model_name=model_name,
+        embedding_mode=embedding_mode,
+        is_recompute=is_recompute,
+    )
+
+    if not skip_search:
+        after_results = run_search(
+            index_path,
+            query,
+            top_k,
+            recompute_embeddings=is_recompute,
+        )
+    else:
+        after_results = None
+    updated_size = index_file_size(index_path)
+
+    return {
+        "initial_size": initial_size,
+        "updated_size": updated_size,
+        "delta": updated_size - initial_size,
+        "before_results": before_results if not skip_search else None,
+        "after_results": after_results if not skip_search else None,
+        "metadata": load_metadata_snapshot(index_path),
+    }
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--initial-files",
+        type=Path,
+        nargs="+",
+        default=DEFAULT_INITIAL_FILES,
+        help="Initial document files (PDF/TXT) used to build the base index",
+    )
+    parser.add_argument(
+        "--index-path",
+        type=Path,
+        default=Path(".leann/examples/leann-demo.leann"),
+        help="Destination index path (default: .leann/examples/leann-demo.leann)",
+    )
+    parser.add_argument(
+        "--initial-count",
+        type=int,
+        default=8,
+        help="Number of chunks to use from the initial documents (default: 8)",
+    )
+    parser.add_argument(
+        "--update-files",
+        type=Path,
+        nargs="*",
+        default=DEFAULT_UPDATE_FILES,
+        help="Additional documents to add during update (PDF/TXT)",
+    )
+    parser.add_argument(
+        "--update-count",
+        type=int,
+        default=4,
+        help="Number of chunks to append from update documents (default: 4)",
+    )
+    parser.add_argument(
+        "--update-text",
+        type=str,
+        default=(
+            "LEANN (Lightweight Embedding ANN) is an indexing toolkit focused on "
+            "recompute-aware HNSW graphs, allowing embeddings to be regenerated "
+            "on demand to keep disk usage minimal."
+        ),
+        help="Fallback text to append if --update-files is omitted",
+    )
+    parser.add_argument(
+        "--top-k",
+        type=int,
+        default=4,
+        help="Number of results to show for each search (default: 4)",
+    )
+    parser.add_argument(
+        "--query",
+        type=str,
+        default=DEFAULT_QUERY,
+        help="Query to run before/after the update",
+    )
+    parser.add_argument(
+        "--embedding-model",
+        type=str,
+        default="sentence-transformers/all-MiniLM-L6-v2",
+        help="Embedding model name",
+    )
+    parser.add_argument(
+        "--embedding-mode",
+        type=str,
+        default="sentence-transformers",
+        choices=["sentence-transformers", "openai", "mlx", "ollama"],
+        help="Embedding backend mode",
+    )
+    parser.add_argument(
+        "--compare-no-recompute",
+        dest="compare_no_recompute",
+        action="store_true",
+        help="Also run a baseline with is_recompute=False and report its index growth.",
+    )
+    parser.add_argument(
+        "--skip-compare-no-recompute",
+        dest="compare_no_recompute",
+        action="store_false",
+        help="Skip building the no-recompute baseline.",
+    )
+    parser.add_argument(
+        "--skip-search",
+        dest="skip_search",
+        action="store_true",
+        help="Skip the search step.",
+    )
+    parser.set_defaults(compare_no_recompute=True)
+    args = parser.parse_args()
+
+    ensure_index_dir(args.index_path)
+    register_project_directory(REPO_ROOT)
+
+    initial_chunks = load_chunks_from_files(list(args.initial_files))
+    if not initial_chunks:
+        raise ValueError("No text chunks extracted from the initial files.")
+
+    initial = initial_chunks[: args.initial_count]
+    if not initial:
+        raise ValueError("Initial chunk set is empty after applying --initial-count.")
+
+    if args.update_files:
+        update_chunks = load_chunks_from_files(list(args.update_files))
+        if not update_chunks:
+            raise ValueError("No text chunks extracted from the update files.")
+        to_add = update_chunks[: args.update_count]
+    else:
+        if not args.update_text:
+            raise ValueError("Provide --update-files or --update-text for the update step.")
+        to_add = [args.update_text]
+    if not to_add:
+        raise ValueError("Update chunk set is empty after applying --update-count.")
+
+    recompute_stats = run_workflow(
+        label="recompute",
+        index_path=args.index_path,
+        initial_paragraphs=initial,
+        update_paragraphs=to_add,
+        model_name=args.embedding_model,
+        embedding_mode=args.embedding_mode,
+        is_recompute=True,
+        query=args.query,
+        top_k=args.top_k,
+        skip_search=args.skip_search,
+    )
+
+    if not args.skip_search:
+        print_results("initial search", recompute_stats["before_results"])
+    if not args.skip_search:
+        print_results("after update", recompute_stats["after_results"])
+    print(
+        f"\n[recompute] Index file size change: {recompute_stats['initial_size']} -> {recompute_stats['updated_size']} bytes"
+        f" (Δ {recompute_stats['delta']})"
+    )
+
+    if recompute_stats["metadata"]:
+        meta_view = {k: recompute_stats["metadata"].get(k) for k in ("is_compact", "is_pruned")}
+        print("[recompute] metadata snapshot:")
+        print(json.dumps(meta_view, indent=2))
+
+    if args.compare_no_recompute:
+        baseline_path = (
+            args.index_path.parent / f"{args.index_path.stem}-norecompute{args.index_path.suffix}"
+        )
+        baseline_stats = run_workflow(
+            label="no-recompute",
+            index_path=baseline_path,
+            initial_paragraphs=initial,
+            update_paragraphs=to_add,
+            model_name=args.embedding_model,
+            embedding_mode=args.embedding_mode,
+            is_recompute=False,
+            query=args.query,
+            top_k=args.top_k,
+            skip_search=args.skip_search,
+        )
+
+        print(
+            f"\n[no-recompute] Index file size change: {baseline_stats['initial_size']} -> {baseline_stats['updated_size']} bytes"
+            f" (Δ {baseline_stats['delta']})"
+        )
+
+        after_texts = (
+            [res.text for res in recompute_stats["after_results"]] if not args.skip_search else None
+        )
+        baseline_after_texts = (
+            [res.text for res in baseline_stats["after_results"]] if not args.skip_search else None
+        )
+        if after_texts == baseline_after_texts:
+            print(
+                "[no-recompute] Search results match recompute baseline; see above for the shared output."
+            )
+        else:
+            print("[no-recompute] WARNING: search results differ from recompute baseline.")
+
+        if baseline_stats["metadata"]:
+            meta_view = {k: baseline_stats["metadata"].get(k) for k in ("is_compact", "is_pruned")}
+            print("[no-recompute] metadata snapshot:")
+            print(json.dumps(meta_view, indent=2))
+
+
+if __name__ == "__main__":
+    main()

examples/email_data/LEANN_email_reader.py

@@ -1,122 +0,0 @@
-import os
-import email
-from pathlib import Path
-from typing import List, Any
-from llama_index.core import Document
-from llama_index.core.readers.base import BaseReader
-
-def find_all_messages_directories(root: str = None) -> List[Path]:
-    """
-    Recursively find all 'Messages' directories under the given root.
-    Returns a list of Path objects.
-    """
-    if root is None:
-        # Auto-detect user's mail path
-        home_dir = os.path.expanduser("~")
-        root = os.path.join(home_dir, "Library", "Mail")
-    
-    messages_dirs = []
-    for dirpath, dirnames, filenames in os.walk(root):
-        if os.path.basename(dirpath) == "Messages":
-            messages_dirs.append(Path(dirpath))
-    return messages_dirs
-
-class EmlxReader(BaseReader):
-    """
-    Apple Mail .emlx file reader with embedded metadata.
-    
-    Reads individual .emlx files from Apple Mail's storage format.
-    """
-    
-    def __init__(self, include_html: bool = False) -> None:
-        """
-        Initialize.
-        
-        Args:
-            include_html: Whether to include HTML content in the email body (default: False)
-        """
-        self.include_html = include_html
-    
-    def load_data(self, input_dir: str, **load_kwargs: Any) -> List[Document]:
-        """
-        Load data from the input directory containing .emlx files.
-        
-        Args:
-            input_dir: Directory containing .emlx files
-            **load_kwargs:
-                max_count (int): Maximum amount of messages to read.
-        """
-        docs: List[Document] = []
-        max_count = load_kwargs.get('max_count', 1000)
-        count = 0
-        
-        # Walk through the directory recursively
-        for dirpath, dirnames, filenames in os.walk(input_dir):
-            # Skip hidden directories
-            dirnames[:] = [d for d in dirnames if not d.startswith(".")]
-            
-            for filename in filenames:
-                if count >= max_count:
-                    break
-                    
-                if filename.endswith(".emlx"):
-                    filepath = os.path.join(dirpath, filename)
-                    try:
-                        # Read the .emlx file
-                        with open(filepath, 'r', encoding='utf-8', errors='ignore') as f:
-                            content = f.read()
-                        
-                        # .emlx files have a length prefix followed by the email content
-                        # The first line contains the length, followed by the email
-                        lines = content.split('\n', 1)
-                        if len(lines) >= 2:
-                            email_content = lines[1]
-                            
-                            # Parse the email using Python's email module
-                            try:
-                                msg = email.message_from_string(email_content)
-                                
-                                # Extract email metadata
-                                subject = msg.get('Subject', 'No Subject')
-                                from_addr = msg.get('From', 'Unknown')
-                                to_addr = msg.get('To', 'Unknown')
-                                date = msg.get('Date', 'Unknown')
-                                
-                                # Extract email body
-                                body = ""
-                                if msg.is_multipart():
-                                    for part in msg.walk():
-                                        if part.get_content_type() == "text/plain" or part.get_content_type() == "text/html":
-                                            if part.get_content_type() == "text/html" and not self.include_html:
-                                                continue
-                                            body += part.get_payload(decode=True).decode('utf-8', errors='ignore')
-                                            # break
-                                else:
-                                    body = msg.get_payload(decode=True).decode('utf-8', errors='ignore')
-                                
-                                # Create document content with metadata embedded in text
-                                doc_content = f"""
-[File]: {filename}
-[From]: {from_addr}
-[To]: {to_addr}
-[Subject]: {subject}
-[Date]: {date}
-[EMAIL BODY Start]:
-{body}
-"""
-                                
-                                # No separate metadata - everything is in the text
-                                doc = Document(text=doc_content, metadata={})
-                                docs.append(doc)
-                                count += 1
-                                
-                            except Exception as e:
-                                print(f"Error parsing email from {filepath}: {e}")
-                                continue
-                                
-                    except Exception as e:
-                        print(f"Error reading file {filepath}: {e}")
-                        continue
-        
-        print(f"Loaded {len(docs)} email documents")
-        return docs 

examples/google_history_reader_leann.py

@@ -1,285 +0,0 @@
-import os
-import asyncio
-import argparse
-try:
-    import dotenv
-    dotenv.load_dotenv()
-except ModuleNotFoundError:
-    # python-dotenv is not installed; skip loading environment variables
-    dotenv = None
-from pathlib import Path
-from typing import List, Any
-from leann.api import LeannBuilder, LeannSearcher, LeannChat
-from llama_index.core.node_parser import SentenceSplitter
-
-# dotenv.load_dotenv()  # handled above if python-dotenv is available
-
-# Default Chrome profile path
-DEFAULT_CHROME_PROFILE = os.path.expanduser("~/Library/Application Support/Google/Chrome/Default")
-
-def create_leann_index_from_multiple_chrome_profiles(profile_dirs: List[Path], index_path: str = "chrome_history_index.leann", max_count: int = -1):
-    """
-    Create LEANN index from multiple Chrome profile data sources.
-    
-    Args:
-        profile_dirs: List of Path objects pointing to Chrome profile directories
-        index_path: Path to save the LEANN index
-        max_count: Maximum number of history entries to process per profile
-    """
-    print("Creating LEANN index from multiple Chrome profile data sources...")
-    
-    # Load documents using ChromeHistoryReader from history_data
-    from history_data.history import ChromeHistoryReader
-    reader = ChromeHistoryReader()
-    
-    INDEX_DIR = Path(index_path).parent
-    
-    if not INDEX_DIR.exists():
-        print(f"--- Index directory not found, building new index ---")
-        all_documents = []
-        total_processed = 0
-        
-        # Process each Chrome profile directory
-        for i, profile_dir in enumerate(profile_dirs):
-            print(f"\nProcessing Chrome profile {i+1}/{len(profile_dirs)}: {profile_dir}")
-            
-            try:
-                documents = reader.load_data(
-                    chrome_profile_path=str(profile_dir),
-                    max_count=max_count
-                )
-                if documents:
-                    print(f"Loaded {len(documents)} history documents from {profile_dir}")
-                    all_documents.extend(documents)
-                    total_processed += len(documents)
-                    
-                    # Check if we've reached the max count
-                    if max_count > 0 and total_processed >= max_count:
-                        print(f"Reached max count of {max_count} documents")
-                        break
-                else:
-                    print(f"No documents loaded from {profile_dir}")
-            except Exception as e:
-                print(f"Error processing {profile_dir}: {e}")
-                continue
-        
-        if not all_documents:
-            print("No documents loaded from any source. Exiting.")
-            # highlight info that you need to close all chrome browser before running this script and high light the instruction!!
-            print("\033[91mYou need to close or quit all chrome browser before running this script\033[0m")
-            return None
-        
-        print(f"\nTotal loaded {len(all_documents)} history documents from {len(profile_dirs)} profiles")
-        
-        # Create text splitter with 256 chunk size
-        text_splitter = SentenceSplitter(chunk_size=256, chunk_overlap=128)
-        
-        # Convert Documents to text strings and chunk them
-        all_texts = []
-        for doc in all_documents:
-            # Split the document into chunks
-            nodes = text_splitter.get_nodes_from_documents([doc])
-            for node in nodes:
-                text = node.get_content()
-                # text = '[Title] ' + doc.metadata["title"] + '\n' + text
-                all_texts.append(text)
-        
-        print(f"Created {len(all_texts)} text chunks from {len(all_documents)} documents")
-        
-        # Create LEANN index directory
-        print(f"--- Index directory not found, building new index ---")
-        INDEX_DIR.mkdir(exist_ok=True)
-
-        print(f"--- Building new LEANN index ---")
-        
-        print(f"\n[PHASE 1] Building Leann index...")
-
-        # Use HNSW backend for better macOS compatibility
-        builder = LeannBuilder(
-            backend_name="hnsw",
-            embedding_model="facebook/contriever",
-            graph_degree=32, 
-            complexity=64,
-            is_compact=True,
-            is_recompute=True,
-            num_threads=1  # Force single-threaded mode
-        )
-
-        print(f"Adding {len(all_texts)} history chunks to index...")
-        for chunk_text in all_texts:
-            builder.add_text(chunk_text)
-            
-        builder.build_index(index_path)
-        print(f"\nLEANN index built at {index_path}!")
-    else:
-        print(f"--- Using existing index at {INDEX_DIR} ---")
-    
-    return index_path
-
-def create_leann_index(profile_path: str = None, index_path: str = "chrome_history_index.leann", max_count: int = 1000):
-    """
-    Create LEANN index from Chrome history data.
-    
-    Args:
-        profile_path: Path to the Chrome profile directory (optional, uses default if None)
-        index_path: Path to save the LEANN index
-        max_count: Maximum number of history entries to process
-    """
-    print("Creating LEANN index from Chrome history data...")
-    INDEX_DIR = Path(index_path).parent
-    
-    if not INDEX_DIR.exists():
-        print(f"--- Index directory not found, building new index ---")
-        INDEX_DIR.mkdir(exist_ok=True)
-
-        print(f"--- Building new LEANN index ---")
-        
-        print(f"\n[PHASE 1] Building Leann index...")
-
-        # Load documents using ChromeHistoryReader from history_data
-        from history_data.history import ChromeHistoryReader
-        reader = ChromeHistoryReader()
-        
-        documents = reader.load_data(
-            chrome_profile_path=profile_path,
-            max_count=max_count
-        )
-        
-        if not documents:
-            print("No documents loaded. Exiting.")
-            return None
-        
-        print(f"Loaded {len(documents)} history documents")
-        
-        # Create text splitter with 256 chunk size
-        text_splitter = SentenceSplitter(chunk_size=256, chunk_overlap=25)
-        
-        # Convert Documents to text strings and chunk them
-        all_texts = []
-        for doc in documents:
-            # Split the document into chunks
-            nodes = text_splitter.get_nodes_from_documents([doc])
-            for node in nodes:
-                all_texts.append(node.get_content())
-        
-        print(f"Created {len(all_texts)} text chunks from {len(documents)} documents")
-        
-        # Create LEANN index directory
-        print(f"--- Index directory not found, building new index ---")
-        INDEX_DIR.mkdir(exist_ok=True)
-
-        print(f"--- Building new LEANN index ---")
-        
-        print(f"\n[PHASE 1] Building Leann index...")
-
-        # Use HNSW backend for better macOS compatibility
-        builder = LeannBuilder(
-            backend_name="hnsw",
-            embedding_model="facebook/contriever",
-            graph_degree=32, 
-            complexity=64,
-            is_compact=True,
-            is_recompute=True,
-            num_threads=1  # Force single-threaded mode
-        )
-
-        print(f"Adding {len(all_texts)} history chunks to index...")
-        for chunk_text in all_texts:
-            builder.add_text(chunk_text)
-            
-        builder.build_index(index_path)
-        print(f"\nLEANN index built at {index_path}!")
-    else:
-        print(f"--- Using existing index at {INDEX_DIR} ---")
-    
-    return index_path
-
-async def query_leann_index(index_path: str, query: str):
-    """
-    Query the LEANN index.
-    
-    Args:
-        index_path: Path to the LEANN index
-        query: The query string
-    """
-    print(f"\n[PHASE 2] Starting Leann chat session...")
-    chat = LeannChat(index_path=index_path)
-    
-    print(f"You: {query}")
-    chat_response = chat.ask(
-        query, 
-        top_k=10, 
-        recompute_beighbor_embeddings=True,
-        complexity=32,
-        beam_width=1,
-        llm_config={
-            "type": "openai",
-            "model": "gpt-4o",
-            "api_key": os.getenv("OPENAI_API_KEY"),
-        },
-        llm_kwargs={
-            "temperature": 0.0,
-            "max_tokens": 1000
-        }
-    )
-    print(f"Leann: {chat_response}")
-
-async def main():
-    # Parse command line arguments
-    parser = argparse.ArgumentParser(description='LEANN Chrome History Reader - Create and query browser history index')
-    parser.add_argument('--chrome-profile', type=str, default=DEFAULT_CHROME_PROFILE,
-                       help=f'Path to Chrome profile directory (default: {DEFAULT_CHROME_PROFILE}), usually you dont need to change this')
-    parser.add_argument('--index-dir', type=str, default="./all_google_new",
-                       help='Directory to store the LEANN index (default: ./chrome_history_index_leann_test)')
-    parser.add_argument('--max-entries', type=int, default=1000,
-                       help='Maximum number of history entries to process (default: 1000)')
-    parser.add_argument('--query', type=str, default=None,
-                       help='Single query to run (default: runs example queries)')
-    parser.add_argument('--auto-find-profiles', action='store_true', default=True,
-                       help='Automatically find all Chrome profiles (default: True)')
-    
-    args = parser.parse_args()
-    
-    INDEX_DIR = Path(args.index_dir)
-    INDEX_PATH = str(INDEX_DIR / "chrome_history.leann")
-    
-    print(f"Using Chrome profile: {args.chrome_profile}")
-    print(f"Index directory: {INDEX_DIR}")
-    print(f"Max entries: {args.max_entries}")
-    
-    # Find Chrome profile directories
-    from history_data.history import ChromeHistoryReader
-    
-    if args.auto_find_profiles:
-        profile_dirs = ChromeHistoryReader.find_chrome_profiles()
-        if not profile_dirs:
-            print("No Chrome profiles found automatically. Exiting.")
-            return
-    else:
-        # Use single specified profile
-        profile_path = Path(args.chrome_profile)
-        if not profile_path.exists():
-            print(f"Chrome profile not found: {profile_path}")
-            return
-        profile_dirs = [profile_path]
-    
-    # Create or load the LEANN index from all sources
-    index_path = create_leann_index_from_multiple_chrome_profiles(profile_dirs, INDEX_PATH, args.max_entries)
-    
-    if index_path:
-        if args.query:
-            # Run single query
-            await query_leann_index(index_path, args.query)
-        else:
-            # Example queries
-            queries = [
-                "What websites did I visit about machine learning?",
-                "Find my search history about programming"
-            ]
-            
-            for query in queries:
-                print("\n" + "="*60)
-                await query_leann_index(index_path, query)
-
-if __name__ == "__main__":
-    asyncio.run(main()) 

examples/grep_search_example.py

@@ -0,0 +1,35 @@
+"""
+Grep Search Example
+
+Shows how to use grep-based text search instead of semantic search.
+Useful when you need exact text matches rather than meaning-based results.
+"""
+
+from leann import LeannSearcher
+
+# Load your index
+searcher = LeannSearcher("my-documents.leann")
+
+# Regular semantic search
+print("=== Semantic Search ===")
+results = searcher.search("machine learning algorithms", top_k=3)
+for result in results:
+    print(f"Score: {result.score:.3f}")
+    print(f"Text: {result.text[:80]}...")
+    print()
+
+# Grep-based search for exact text matches
+print("=== Grep Search ===")
+results = searcher.search("def train_model", top_k=3, use_grep=True)
+for result in results:
+    print(f"Score: {result.score}")
+    print(f"Text: {result.text[:80]}...")
+    print()
+
+# Find specific error messages
+error_results = searcher.search("FileNotFoundError", use_grep=True)
+print(f"Found {len(error_results)} files mentioning FileNotFoundError")
+
+# Search for function definitions
+func_results = searcher.search("class SearchResult", use_grep=True, top_k=5)
+print(f"Found {len(func_results)} class definitions")