feat: implement smart memory configuration for DiskANN

- Add intelligent memory calculation based on data size and system specs
- search_memory_maximum: 1/10 of embedding size (controls PQ compression)
- build_memory_maximum: 50% of available RAM (controls sharding)
- Provides optimal balance between performance and memory usage
- Automatic fallback to default values if parameters are explicitly provided

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

@@ -0,0 +1,11 @@
+name: CI
+
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+
+jobs:
+  build:
+    uses: ./.github/workflows/build-reusable.yml

.github/workflows/build-reusable.yml

@@ -0,0 +1,251 @@
+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 }}
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+
+      - name: Install ruff
+        run: |
+          uv tool install ruff
+
+      - name: Run ruff check
+        run: |
+          ruff check .
+
+      - name: Run ruff format check
+        run: |
+          ruff format --check .
+
+  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'
+          - os: macos-latest
+            python: '3.9'
+          - os: macos-latest
+            python: '3.10'
+          - os: macos-latest
+            python: '3.11'
+          - os: macos-latest
+            python: '3.12'
+          - os: macos-latest
+            python: '3.13'
+    runs-on: ${{ matrix.os }}
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+          submodules: recursive
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+
+      - 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 protobuf-compiler libzmq3-dev \
+            pkg-config libopenblas-dev patchelf libabsl-dev libaio-dev libprotobuf-dev
+
+          # Install Intel MKL for DiskANN
+          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" >> $GITHUB_ENV
+          echo "LD_LIBRARY_PATH=/opt/intel/oneapi/mkl/latest/lib/intel64:$LD_LIBRARY_PATH" >> $GITHUB_ENV
+
+      - name: Install system dependencies (macOS)
+        if: runner.os == 'macOS'
+        run: |
+          # Don't install LLVM, use system clang for better compatibility
+          brew install libomp boost protobuf zeromq
+
+      - name: Install build dependencies
+        run: |
+          uv pip install --system scikit-build-core numpy swig Cython pybind11
+          if [[ "$RUNNER_OS" == "Linux" ]]; then
+            uv pip install --system auditwheel
+          else
+            uv pip install --system delocate
+          fi
+
+      - name: Build packages
+        run: |
+          # Build core (platform independent)
+          if [[ "${{ matrix.os }}" == ubuntu-* ]]; then
+            cd packages/leann-core
+            uv build
+            cd ../..
+          fi
+
+          # Build HNSW backend
+          cd packages/leann-backend-hnsw
+          if [ "${{ matrix.os }}" == "macos-latest" ]; then
+            # Use system clang instead of homebrew LLVM for better compatibility
+            export CC=clang
+            export CXX=clang++
+            export MACOSX_DEPLOYMENT_TARGET=11.0
+            uv build --wheel --python python
+          else
+            uv build --wheel --python python
+          fi
+          cd ../..
+
+          # Build DiskANN backend
+          cd packages/leann-backend-diskann
+          if [ "${{ matrix.os }}" == "macos-latest" ]; then
+            # Use system clang instead of homebrew LLVM for better compatibility
+            export CC=clang
+            export CXX=clang++
+            # DiskANN requires macOS 13.3+ for sgesdd_ LAPACK function
+            export MACOSX_DEPLOYMENT_TARGET=13.3
+            uv build --wheel --python python
+          else
+            uv build --wheel --python python
+          fi
+          cd ../..
+
+          # Build meta package (platform independent)
+          if [[ "${{ matrix.os }}" == ubuntu-* ]]; then
+            cd packages/leann
+            uv build
+            cd ../..
+          fi
+
+      - 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: |
+          # Repair HNSW wheel
+          cd packages/leann-backend-hnsw
+          if [ -d dist ]; then
+            delocate-wheel -w dist_repaired -v dist/*.whl
+            rm -rf dist
+            mv dist_repaired dist
+          fi
+          cd ../..
+
+          # Repair DiskANN wheel
+          cd packages/leann-backend-diskann
+          if [ -d dist ]; then
+            delocate-wheel -w dist_repaired -v 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 a virtual environment
+          uv venv
+          source .venv/bin/activate || source .venv/Scripts/activate
+
+          # Install the built wheels
+          # Use --find-links to let uv choose the correct wheel for the platform
+          if [[ "${{ matrix.os }}" == ubuntu-* ]]; then
+            uv pip install leann-core --find-links packages/leann-core/dist
+            uv pip install leann --find-links packages/leann/dist
+          fi
+          uv pip install leann-backend-hnsw --find-links packages/leann-backend-hnsw/dist
+          uv pip install leann-backend-diskann --find-links packages/leann-backend-diskann/dist
+
+          # Install test dependencies using extras
+          uv pip install -e ".[test]"
+
+      - name: Run tests with pytest
+        env:
+          CI: true  # Mark as CI environment to skip memory-intensive tests
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          HF_HUB_DISABLE_SYMLINKS: 1
+          TOKENIZERS_PARALLELISM: false
+          PYTORCH_ENABLE_MPS_FALLBACK: 0  # Disable MPS on macOS CI to avoid memory issues
+          OMP_NUM_THREADS: 1  # Disable OpenMP parallelism to avoid libomp crashes
+          MKL_NUM_THREADS: 1  # Single thread for MKL operations
+        run: |
+          # Activate virtual environment
+          source .venv/bin/activate || source .venv/Scripts/activate
+
+          # Run all tests
+          pytest tests/
+
+      - name: Run sanity checks (optional)
+        run: |
+          # Activate virtual environment
+          source .venv/bin/activate || source .venv/Scripts/activate
+
+          # Run distance function tests if available
+          if [ -f test/sanity_checks/test_distance_functions.py ]; then
+            echo "Running distance function sanity checks..."
+            python test/sanity_checks/test_distance_functions.py || echo "⚠️ Distance function test failed, continuing..."
+          fi
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: packages-${{ matrix.os }}-py${{ matrix.python }}
+          path: packages/*/dist/

.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

@@ -9,10 +9,9 @@ demo/indices/
 outputs/
 *.pkl
 *.pdf
-*.idx 
+*.idx
 *.map
 .history/
-scripts/
 lm_eval.egg-info/
 demo/experiment_results/**/*.json
 *.jsonl
@@ -84,4 +83,8 @@ test_*.py
 packages/leann-backend-diskann/third_party/DiskANN/_deps/
 
 *.meta.json
-*.passages.json
+*.passages.json
+
+batchtest.py
+tests/__pytest_cache__/
+tests/__pycache__/

.pre-commit-config.yaml

@@ -0,0 +1,16 @@
+repos:
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.5.0
+    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: v0.2.1
+    hooks:
+      - id: ruff
+      - id: ruff-format

.vscode/extensions.json

@@ -1,9 +0,0 @@
-{
-    "recommendations": [
-        "llvm-vs-code-extensions.vscode-clangd",
-        "ms-python.python",
-        "ms-vscode.cmake-tools",
-        "vadimcn.vscode-lldb",
-        "eamodio.gitlens",
-    ]
-}

.vscode/launch.json

@@ -1,283 +0,0 @@
-{
-    // Use IntelliSense to learn about possible attributes.
-    // Hover to view descriptions of existing attributes.
-    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
-    "version": "0.2.0",
-    "configurations": [
-        // new emdedder
-        {
-            "name": "New Embedder",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "demo/main.py",
-            "console": "integratedTerminal",
-            "args": [
-                "--search",
-                "--use-original",
-                "--domain",
-                "dpr",
-                "--nprobe",
-                "5000",
-                "--load",
-                "flat",
-                "--embedder",
-                "intfloat/multilingual-e5-small"
-            ]
-        }
-        //python /home/ubuntu/Power-RAG/faiss/demo/simple_build.py
-        {
-            "name": "main.py",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "demo/main.py",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "--query",
-                "1000",
-                "--load",
-                "bm25"
-            ]
-        },
-        {
-            "name": "Simple Build",
-            "type": "lldb",
-            "request": "launch",
-            "program": "${workspaceFolder}/.venv/bin/python",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "faiss/demo/simple_build.py"
-            ],
-            "env": {
-                "LD_PRELOAD": "/lib/x86_64-linux-gnu/libmkl_core.so:/lib/x86_64-linux-gnu/libmkl_intel_thread.so:/lib/x86_64-linux-gnu/libmkl_intel_lp64.so:/lib/x86_64-linux-gnu/libiomp5.so"
-            }
-        },
-        //# Fix for Intel MKL error
-        //export LD_PRELOAD=/lib/x86_64-linux-gnu/libmkl_core.so:/lib/x86_64-linux-gnu/libmkl_intel_thread.so:/lib/x86_64-linux-gnu/libmkl_intel_lp64.so:/lib/x86_64-linux-gnu/libiomp5.so
-        //python faiss/demo/build_demo.py
-        {
-            "name": "Build Demo",
-            "type": "lldb",
-            "request": "launch",
-            "program": "${workspaceFolder}/.venv/bin/python",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "faiss/demo/build_demo.py"
-            ],
-            "env": {
-                "LD_PRELOAD": "/lib/x86_64-linux-gnu/libmkl_core.so:/lib/x86_64-linux-gnu/libmkl_intel_thread.so:/lib/x86_64-linux-gnu/libmkl_intel_lp64.so:/lib/x86_64-linux-gnu/libiomp5.so"
-            }
-        },
-        {
-            "name": "DiskANN Serve",
-            "type": "lldb",
-            "request": "launch",
-            "program": "${workspaceFolder}/.venv/bin/python",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "demo/main.py",
-                "--mode",
-                "serve",
-                "--engine",
-                "sglang",
-                "--load-indices",
-                "diskann",
-                "--domain",
-                "rpj_wiki",
-                "--lazy-load",
-                "--recompute-beighbor-embeddings",
-                "--port",
-                "8082",
-                "--diskann-search-memory-maximum",
-                "2",
-                "--diskann-graph",
-                "240",
-                "--search-only"
-            ],
-            "env": {
-                "PYTHONPATH": "${workspaceFolder}/faiss_repo/build/faiss/python:$PYTHONPATH"
-            },
-            "preLaunchTask": "CMake: build",
-        },
-        {
-            "name": "DiskANN Serve MAC",
-            "type": "lldb",
-            "request": "launch",
-            "program": "${workspaceFolder}/.venv/bin/python",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "demo/main.py",
-                "--mode",
-                "serve",
-                "--engine",
-                "ollama",
-                "--load-indices",
-                "diskann",
-                "--domain",
-                "rpj_wiki",
-                "--lazy-load",
-                "--recompute-beighbor-embeddings"
-            ],
-            "preLaunchTask": "CMake: build",
-            "env": {
-                "KMP_DUPLICATE_LIB_OK": "TRUE",
-                "OMP_NUM_THREADS": "1",
-                "MKL_NUM_THREADS": "1",
-                "DYLD_INSERT_LIBRARIES": "/Users/ec2-user/Power-RAG/.venv/lib/python3.10/site-packages/torch/lib/libomp.dylib",
-                "KMP_BLOCKTIME": "0"
-            }
-        },
-        {
-            "name": "Python Debugger: Current File with Arguments",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "ric/main_ric.py",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "--config-name",
-                "${input:configSelection}"
-            ],
-            "justMyCode": false
-        },
-        //python ./demo/validate_equivalence.py sglang
-        {
-            "name": "Validate Equivalence",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "demo/validate_equivalence.py",
-            "console": "integratedTerminal",
-            "args": [
-                "sglang"
-            ],
-        },
-        //python demo/retrieval_demo.py --engine sglang  --skip-embeddings --domain dpr --load-indices flat ivf_flat
-        {
-            "name": "Retrieval Demo",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "demo/retrieval_demo.py",
-            "console": "integratedTerminal",
-            "args": [
-                "--engine",
-                "vllm",
-                "--skip-embeddings",
-                "--domain",
-                "dpr",
-                "--load-indices",
-                // "flat",
-                "ivf_flat"
-            ],
-        },
-        //python demo/retrieval_demo.py --engine sglang  --skip-embeddings --domain dpr  --load-indices  diskann --hnsw-M 64 --hnsw-efConstruction 150 --hnsw-efSearch 128  --hnsw-sq-bits 8 
-        {
-            "name": "Retrieval Demo DiskANN",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "demo/retrieval_demo.py",
-            "console": "integratedTerminal",
-            "args": [
-                "--engine",
-                "sglang",
-                "--skip-embeddings",
-                "--domain",
-                "dpr",
-                "--load-indices",
-                "diskann",
-                "--hnsw-M",
-                "64",
-                "--hnsw-efConstruction",
-                "150",
-                "--hnsw-efSearch",
-                "128",
-                "--hnsw-sq-bits",
-                "8"
-            ],
-        },
-        {
-            "name": "Find Probe",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "find_probe.py",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-        },
-        {
-            "name": "Python: Attach",
-            "type": "debugpy",
-            "request": "attach",
-            "processId": "${command:pickProcess}",
-            "justMyCode": true
-        },
-        {
-            "name": "Edge RAG",
-            "type": "lldb",
-            "request": "launch",
-            "program": "${workspaceFolder}/.venv/bin/python",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "edgerag_demo.py"
-            ],
-            "env": {
-                "LD_PRELOAD": "/lib/x86_64-linux-gnu/libiomp5.so /lib/x86_64-linux-gnu/libmkl_core.so /lib/x86_64-linux-gnu/libmkl_intel_lp64.so /lib/x86_64-linux-gnu/libmkl_intel_thread.so",
-                "MKL_NUM_THREADS": "1",
-                "OMP_NUM_THREADS": "1",
-            }
-        },
-        {
-            "name": "Launch Embedding Server",
-            "type": "debugpy",
-            "request": "launch",
-            "program": "demo/embedding_server.py",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "--domain",
-                "rpj_wiki",
-                "--zmq-port",
-                "5556",
-            ]
-        },
-        {
-            "name": "HNSW Serve",
-            "type": "lldb",
-            "request": "launch",
-            "program": "${workspaceFolder}/.venv/bin/python",
-            "console": "integratedTerminal",
-            "cwd": "${workspaceFolder}",
-            "args": [
-                "demo/main.py",
-                "--domain",
-                "rpj_wiki",
-                "--load",
-                "hnsw",
-                "--mode",
-                "serve",
-                "--search",
-                "--skip-pa",
-                "--recompute",
-                "--hnsw-old"
-            ],
-            "env": {
-                "LD_PRELOAD": "/lib/x86_64-linux-gnu/libmkl_core.so:/lib/x86_64-linux-gnu/libmkl_intel_thread.so:/lib/x86_64-linux-gnu/libmkl_intel_lp64.so:/lib/x86_64-linux-gnu/libiomp5.so"
-            }
-        },
-    ],
-    "inputs": [
-        {
-            "id": "configSelection",
-            "type": "pickString",
-            "description": "Select a configuration",
-            "options": [
-                "example_config",
-                "vllm_gritlm"
-            ],
-            "default": "example_config"
-        }
-    ],
-}

.vscode/settings.json

@@ -1,43 +0,0 @@
-{
-    "python.analysis.extraPaths": [
-        "./sglang_repo/python"
-    ],
-    "cmake.sourceDirectory": "${workspaceFolder}/DiskANN",
-    "cmake.configureArgs": [
-        "-DPYBIND=True",
-        "-DUPDATE_EDITABLE_INSTALL=ON",
-    ],
-    "cmake.environment": {
-        "PATH": "/Users/ec2-user/Power-RAG/.venv/bin:${env:PATH}"
-    },
-    "cmake.buildDirectory": "${workspaceFolder}/build",
-    "files.associations": {
-        "*.tcc": "cpp",
-        "deque": "cpp",
-        "string": "cpp",
-        "unordered_map": "cpp",
-        "vector": "cpp",
-        "map": "cpp",
-        "unordered_set": "cpp",
-        "atomic": "cpp",
-        "inplace_vector": "cpp",
-        "*.ipp": "cpp",
-        "forward_list": "cpp",
-        "list": "cpp",
-        "any": "cpp",
-        "system_error": "cpp",
-        "__hash_table": "cpp",
-        "__split_buffer": "cpp",
-        "__tree": "cpp",
-        "ios": "cpp",
-        "set": "cpp",
-        "__string": "cpp",
-        "string_view": "cpp",
-        "ranges": "cpp",
-        "iosfwd": "cpp"
-    },
-    "lldb.displayFormat": "auto",
-    "lldb.showDisassembly": "auto",
-    "lldb.dereferencePointers": true,
-    "lldb.consoleMode": "commands",
-}

.vscode/tasks.json

@@ -1,16 +0,0 @@
-{
-	"version": "2.0.0",
-	"tasks": [
-		{
-			"type": "cmake",
-			"label": "CMake: build",
-			"command": "build",
-			"targets": [
-				"all"
-			],
-			"group": "build",
-			"problemMatcher": [],
-			"detail": "CMake template build task"
-		}
-	]
-}

README.md

@@ -12,11 +12,11 @@
     The smallest vector index in the world. RAG Everything with LEANN!
 </h2>
 
-LEANN is a revolutionary vector database that democratizes personal AI. Transform your laptop into a powerful RAG system that can index and search through millions of documents while using **[97% less storage]** than traditional solutions **without accuracy loss**.
+LEANN is an innovative vector database that democratizes personal AI. Transform your laptop into a powerful RAG system that can index and search through millions of documents while using **97% less storage** than traditional solutions **without accuracy loss**.
 
-LEANN achieves this through *graph-based selective recomputation* with *high-degree preserving pruning*, computing embeddings on-demand instead of storing them all. [Illustration →](#️-architecture--how-it-works) | [Paper →](https://arxiv.org/abs/2506.08276)
+LEANN achieves this through *graph-based selective recomputation* with *high-degree preserving pruning*, computing embeddings on-demand instead of storing them all. [Illustration Fig →](#️-architecture--how-it-works) | [Paper →](https://arxiv.org/abs/2506.08276)
 
-**Ready to RAG Everything?** Transform your laptop into a personal AI assistant that can search your **[file system](#process-any-documents-pdf-txt-md)**, **[emails](#search-your-entire-life)**, **[browser history](#time-machine-for-the-web)**, **[chat history](#wechat-detective)**, or external knowledge bases (i.e., 60M documents) - all on your laptop, with zero cloud costs and complete privacy.
+**Ready to RAG Everything?** Transform your laptop into a personal AI assistant that can search your **[file system](#-personal-data-manager-process-any-documents-pdf-txt-md)**, **[emails](#-your-personal-email-secretary-rag-on-apple-mail)**, **[browser history](#-time-machine-for-the-web-rag-your-entire-browser-history)**, **[chat history](#-wechat-detective-unlock-your-golden-memories)**, or external knowledge bases (i.e., 60M documents) - all on your laptop, with zero cloud costs and complete privacy.
 
 
 
@@ -26,19 +26,52 @@ LEANN achieves this through *graph-based selective recomputation* with *high-deg
   <img src="assets/effects.png" alt="LEANN vs Traditional Vector DB Storage Comparison" width="70%">
 </p>
 
-**The numbers speak for themselves:** Index 60 million Wikipedia articles in just 6GB instead of 201GB. From emails to browser history, everything fits on your laptop. [See detailed benchmarks below ↓](#storage-usage-comparison)
+> **The numbers speak for themselves:** Index 60 million Wikipedia chunks in just 6GB instead of 201GB. From emails to browser history, everything fits on your laptop. [See detailed benchmarks for different applications below ↓](#storage-comparison)
 
-## Why This Matters
 
 🔒 **Privacy:** Your data never leaves your laptop. No OpenAI, no cloud, no "terms of service".
 
 🪶 **Lightweight:** Graph-based recomputation eliminates heavy embedding storage, while smart graph pruning and CSR format minimize graph storage overhead. Always less storage, less memory usage!
 
+📦 **Portable:** Transfer your entire knowledge base between devices (even with others) with minimal cost - your personal AI memory travels with you.
+
 📈 **Scalability:** Handle messy personal data that would crash traditional vector DBs, easily managing your growing personalized data and agent generated memory!
 
 ✨ **No Accuracy Loss:** Maintain the same search quality as heavyweight solutions while using 97% less storage.
 
-## Quick Start in 1 minute
+## Installation
+
+<details>
+<summary><strong>📦 Prerequisites: Install uv (if you don't have it)</strong></summary>
+
+Install uv first if you don't have it:
+
+```bash
+curl -LsSf https://astral.sh/uv/install.sh | sh
+```
+
+📖 [Detailed uv installation methods →](https://docs.astral.sh/uv/getting-started/installation/#installation-methods)
+
+</details>
+
+
+LEANN provides two installation methods: **pip install** (quick and easy) and **build from source** (recommended for development).
+
+
+
+### 🚀 Quick Install (Recommended for most users)
+
+Clone the repository to access all examples and install LEANN from [PyPI](https://pypi.org/project/leann/) to run them immediately:
+
+```bash
+git clone git@github.com:yichuan-w/LEANN.git leann
+cd leann
+uv venv
+source .venv/bin/activate
+uv pip install leann
+```
+
+### 🔧 Build from Source (Recommended for development)
 
 ```bash
 git clone git@github.com:yichuan-w/LEANN.git leann
@@ -48,33 +81,68 @@ git submodule update --init --recursive
 
 **macOS:**
 ```bash
-brew install llvm libomp boost protobuf zeromq
-export CC=$(brew --prefix llvm)/bin/clang
-export CXX=$(brew --prefix llvm)/bin/clang++
-
-# Install with HNSW backend (default, recommended for most users)
-uv sync
-
-# Or add DiskANN backend if you want to test more options
-uv sync --extra diskann
+brew install llvm libomp boost protobuf zeromq pkgconf
+CC=$(brew --prefix llvm)/bin/clang CXX=$(brew --prefix llvm)/bin/clang++ uv sync
 ```
 
-**Linux (Ubuntu/Debian):**
+**Linux:**
 ```bash
 sudo apt-get install libomp-dev libboost-all-dev protobuf-compiler libabsl-dev libmkl-full-dev libaio-dev libzmq3-dev
-
-# Install with HNSW backend (default, recommended for most users)
 uv sync
-
-# Or add DiskANN backend if you want to test more options
-uv sync --extra diskann
 ```
 
-**Ollama Setup (Optional for Local LLM):**
 
-*We support both hf-transformers and Ollama for local LLMs. Ollama is recommended for faster performance.*
 
-*macOS:*
+
+## Quick Start
+
+Our declarative API makes RAG as easy as writing a config file.
+
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yichuan-w/LEANN/blob/main/demo.ipynb) [Try in this ipynb file →](demo.ipynb)
+
+```python
+from leann import LeannBuilder, LeannSearcher, LeannChat
+from pathlib import Path
+INDEX_PATH = str(Path("./").resolve() / "demo.leann")
+
+# Build an index
+builder = LeannBuilder(backend_name="hnsw")
+builder.add_text("LEANN saves 97% storage compared to traditional vector databases.")
+builder.add_text("Tung Tung Tung Sahur called—they need their banana‑crocodile hybrid back")
+builder.build_index(INDEX_PATH)
+
+# Search
+searcher = LeannSearcher(INDEX_PATH)
+results = searcher.search("fantastical AI-generated creatures", top_k=1)
+
+# Chat with your data
+chat = LeannChat(INDEX_PATH, llm_config={"type": "hf", "model": "Qwen/Qwen3-0.6B"})
+response = chat.ask("How much storage does LEANN save?", top_k=1)
+```
+
+## RAG on Everything!
+
+LEANN supports RAG on various data sources including documents (.pdf, .txt, .md), Apple Mail, Google Search History, WeChat, and more.
+
+
+> **Generation Model Setup**
+> LEANN supports multiple LLM providers for text generation (OpenAI API, HuggingFace, Ollama).
+
+<details>
+<summary><strong>🔑 OpenAI API Setup (Default)</strong></summary>
+
+Set your OpenAI API key as an environment variable:
+
+```bash
+export OPENAI_API_KEY="your-api-key-here"
+```
+
+</details>
+
+<details>
+<summary><strong>🔧 Ollama Setup (Recommended for full privacy)</strong></summary>
+
+**macOS:**
 
 First, [download Ollama for macOS](https://ollama.com/download/mac).
 
@@ -83,7 +151,8 @@ First, [download Ollama for macOS](https://ollama.com/download/mac).
 ollama pull llama3.2:1b
 ```
 
-*Linux:*
+**Linux:**
+
 ```bash
 # Install Ollama
 curl -fsSL https://ollama.ai/install.sh | sh
@@ -95,65 +164,56 @@ ollama serve &
 ollama pull llama3.2:1b
 ```
 
-You can also replace `llama3.2:1b` to `deepseek-r1:1.5b` or `qwen3:4b` for better performance but higher memory usage.
+</details>
 
-## Dead Simple API
+### 📄 Personal Data Manager: Process Any Documents (.pdf, .txt, .md)!
 
-Just 3 lines of code. Our declarative API makes RAG as easy as writing a config file:
+Ask questions directly about your personal PDFs, documents, and any directory containing your files!
 
-```python
-from leann.api import LeannBuilder, LeannSearcher
+<p align="center">
+  <img src="videos/paper_clear.gif" alt="LEANN Document Search Demo" width="600">
+</p>
 
-# 1. Build index (no embeddings stored!)
-builder = LeannBuilder(backend_name="hnsw")
-builder.add_text("C# is a powerful programming language")
-builder.add_text("Python is a powerful programming language")
-builder.add_text("Machine learning transforms industries")  
-builder.add_text("Neural networks process complex data")
-builder.add_text("Leann is a great storage saving engine for RAG on your macbook")
-builder.build_index("knowledge.leann")
-
-# 2. Search with real-time embeddings
-searcher = LeannSearcher("knowledge.leann")
-results = searcher.search("C++ programming languages", top_k=2, recompute_beighbor_embeddings=True)
-print(results)
-```
-
-**That's it.** No cloud setup, no API keys, no "fine-tuning". Just your data, your questions, your laptop.
-
-[Try the interactive demo →](demo.ipynb)
-
-## Wild Things You Can Do
-
-LEANN supports RAGing a lot of data sources, like .pdf, .txt, .md, and also supports RAGing your WeChat, Google Search History, and more.
-
-### Process Any Documents (.pdf, .txt, .md)
-
-Above we showed the Python API, while this CLI script demonstrates the same concepts while directly processing PDFs and documents.
+The example below asks a question about summarizing two papers (uses default data in `examples/data`) and this is the easiest example to run here:
 
 ```bash
-# Drop your PDFs, .txt, .md files into examples/data/
-uv run ./examples/main_cli_example.py
-
-# Or use python directly
 source .venv/bin/activate
 python ./examples/main_cli_example.py
 ```
 
-Uses Ollama `qwen3:8b` by default. For other models: `--llm openai --model gpt-4o` (requires `OPENAI_API_KEY` environment variable) or `--llm hf --model Qwen/Qwen3-4B`.
+<details>
+<summary><strong>📋 Click to expand: User Configurable Arguments</strong></summary>
 
-**Works with any text format** - research papers, personal notes, presentations. Built with LlamaIndex for document parsing.
-
-### Search Your Entire Life
 ```bash
-python examples/mail_reader_leann.py
-# "What did my boss say about the Christmas party last year?"
-# "Find all emails from my mom about birthday plans"
+# Use custom index directory
+python examples/main_cli_example.py --index-dir "./my_custom_index"
+
+# Use custom data directory
+python examples/main_cli_example.py --data-dir "./my_documents"
+
+# Ask a specific question
+python examples/main_cli_example.py --query "What are the main findings in these papers?"
 ```
-**90K emails → 14MB.** Finally, search your email like you search Google.
+
+</details>
+
+### 📧 Your Personal Email Secretary: RAG on Apple Mail!
+
+> **Note:** The examples below currently support macOS only. Windows support coming soon.
+
+
+<p align="center">
+  <img src="videos/mail_clear.gif" alt="LEANN Email Search Demo" width="600">
+</p>
+
+**Note:** You need to grant full disk access to your terminal/VS Code in System Preferences → Privacy & Security → Full Disk Access.
+```bash
+python examples/mail_reader_leann.py --query "What's the food I ordered by DoorDash or Uber Eats mostly?"
+```
+**780K email chunks → 78MB storage.** Finally, search your email like you search Google.
 
 <details>
-<summary><strong>📋 Click to expand: Command Examples</strong></summary>
+<summary><strong>📋 Click to expand: User Configurable Arguments</strong></summary>
 
 ```bash
 # Use default mail path (works for most macOS setups)
@@ -183,16 +243,19 @@ Once the index is built, you can ask questions like:
 - "Show me emails about travel expenses"
 </details>
 
-### Time Machine for the Web  
+### 🔍 Time Machine for the Web: RAG Your Entire Chrome Browser History!
+
+<p align="center">
+  <img src="videos/google_clear.gif" alt="LEANN Browser History Search Demo" width="600">
+</p>
+
 ```bash
-python examples/google_history_reader_leann.py
-# "What was that AI paper I read last month?"
-# "Show me all the cooking videos I watched"
+python examples/google_history_reader_leann.py --query "Tell me my browser history about machine learning?"
 ```
-**38K browser entries → 6MB.** Your browser history becomes your personal search engine.
+**38K browser entries → 6MB storage.** Your browser history becomes your personal search engine.
 
 <details>
-<summary><strong>📋 Click to expand: Command Examples</strong></summary>
+<summary><strong>📋 Click to expand: User Configurable Arguments</strong></summary>
 
 ```bash
 # Use default Chrome profile (auto-finds all profiles)
@@ -238,13 +301,17 @@ Once the index is built, you can ask questions like:
 
 </details>
 
-### WeChat Detective
+### 💬 WeChat Detective: Unlock Your Golden Memories!
+
+<p align="center">
+  <img src="videos/wechat_clear.gif" alt="LEANN WeChat Search Demo" width="600">
+</p>
 
 ```bash
-python examples/wechat_history_reader_leann.py
-# "Show me all group chats about weekend plans"
+python examples/wechat_history_reader_leann.py --query "Show me all group chats about weekend plans"
 ```
-**400K messages → 64MB.** Search years of chat history in any language.
+**400K messages → 64MB storage** Search years of chat history in any language.
+
 
 <details>
 <summary><strong>🔧 Click to expand: Installation Requirements</strong></summary>
@@ -255,11 +322,17 @@ First, you need to install the WeChat exporter:
 sudo packages/wechat-exporter/wechattweak-cli install
 ```
 
-**Troubleshooting**: If you encounter installation issues, check the [WeChatTweak-CLI issues page](https://github.com/sunnyyoung/WeChatTweak-CLI/issues/41).
+**Troubleshooting:**
+- **Installation issues**: Check the [WeChatTweak-CLI issues page](https://github.com/sunnyyoung/WeChatTweak-CLI/issues/41)
+- **Export errors**: If you encounter the error below, try restarting WeChat
+```
+Failed to export WeChat data. Please ensure WeChat is running and WeChatTweak is installed.
+Failed to find or export WeChat data. Exiting.
+```
 </details>
 
 <details>
-<summary><strong>📋 Click to expand: Command Examples</strong></summary>
+<summary><strong>📋 Click to expand: User Configurable Arguments</strong></summary>
 
 ```bash
 # Use default settings (recommended for first run)
@@ -290,6 +363,73 @@ Once the index is built, you can ask questions like:
 </details>
 
 
+
+## 🖥️ Command Line Interface
+
+LEANN includes a powerful CLI for document processing and search. Perfect for quick document indexing and interactive chat.
+
+```bash
+# Build an index from documents
+leann build my-docs --docs ./documents
+
+# Search your documents
+leann search my-docs "machine learning concepts"
+
+# Interactive chat with your documents
+leann ask my-docs --interactive
+
+# List all your indexes
+leann list
+```
+
+**Key CLI features:**
+- Auto-detects document formats (PDF, TXT, MD, DOCX)
+- Smart text chunking with overlap
+- Multiple LLM providers (Ollama, OpenAI, HuggingFace)
+- Organized index storage in `~/.leann/indexes/`
+- Support for advanced search parameters
+
+<details>
+<summary><strong>📋 Click to expand: Complete CLI Reference</strong></summary>
+
+**Build Command:**
+```bash
+leann build INDEX_NAME --docs DIRECTORY [OPTIONS]
+
+Options:
+  --backend {hnsw,diskann}     Backend to use (default: hnsw)
+  --embedding-model MODEL      Embedding model (default: facebook/contriever)
+  --graph-degree N            Graph degree (default: 32)
+  --complexity N              Build complexity (default: 64)
+  --force                     Force rebuild existing index
+  --compact                   Use compact storage (default: true)
+  --recompute                 Enable recomputation (default: true)
+```
+
+**Search Command:**
+```bash
+leann search INDEX_NAME QUERY [OPTIONS]
+
+Options:
+  --top-k N                   Number of results (default: 5)
+  --complexity N              Search complexity (default: 64)
+  --recompute-embeddings      Use recomputation for highest accuracy
+  --pruning-strategy {global,local,proportional}
+```
+
+**Ask Command:**
+```bash
+leann ask INDEX_NAME [OPTIONS]
+
+Options:
+  --llm {ollama,openai,hf}    LLM provider (default: ollama)
+  --model MODEL               Model name (default: qwen3:8b)
+  --interactive              Interactive chat mode
+  --top-k N                  Retrieval count (default: 20)
+```
+
+</details>
+
 ## 🏗️ Architecture & How It Works
 
 <p align="center">
@@ -300,7 +440,7 @@ Once the index is built, you can ask questions like:
 
 **Core techniques:**
 - **Graph-based selective recomputation:** Only compute embeddings for nodes in the search path
-- **High-degree preserving pruning:** Keep important "hub" nodes while removing redundant connections  
+- **High-degree preserving pruning:** Keep important "hub" nodes while removing redundant connections
 - **Dynamic batching:** Efficiently batch embedding computations for GPU utilization
 - **Two-level search:** Smart graph traversal that prioritizes promising nodes
 
@@ -308,18 +448,17 @@ Once the index is built, you can ask questions like:
 
 ## Benchmarks
 
-Run the comparison yourself:
-```bash
-python examples/compare_faiss_vs_leann.py
-```
 
-| System | Storage | 
-|--------|---------|
-| FAISS HNSW | 5.5 MB |
-| LEANN | 0.5 MB |
-| **Savings** | **91%** |
+📊 **[Simple Example: Compare LEANN vs FAISS →](examples/compare_faiss_vs_leann.py)**
+### Storage Comparison
+
+| System | DPR (2.1M) | Wiki (60M) | Chat (400K) | Email (780K) | Browser (38K) |
+|--------|-------------|------------|-------------|--------------|---------------|
+| Traditional vector database (e.g., FAISS) | 3.8 GB      | 201 GB     | 1.8 GB     | 2.4 GB      | 130 MB        |
+| LEANN  | 324 MB      | 6 GB       | 64 MB       | 79 MB       | 6.4 MB        |
+| Savings| 91%         | 97%        | 97%         | 97%         | 95%           |
+
 
-Same dataset, same hardware, same embedding model. LEANN just works better.
 
 ## Reproduce Our Results
 
@@ -329,33 +468,7 @@ python examples/run_evaluation.py data/indices/dpr/dpr_diskann      # DPR datase
 python examples/run_evaluation.py data/indices/rpj_wiki/rpj_wiki.index  # Wikipedia
 ```
 
-The evaluation script downloads data automatically on first run.
-
-### Storage Usage Comparison
-
-| System                | DPR (2.1M chunks) | RPJ-wiki (60M chunks) | Chat history (400K messages) | Apple emails (90K messages chunks) |Google Search History (38K entries)
-|-----------------------|------------------|------------------------|-----------------------------|------------------------------|------------------------------|
-| Traditional Vector DB(FAISS) | 3.8 GB           | 201 GB                 | 1.8G                     | 305.8 MB                     |130.4 MB                     |
-| **LEANN**             | **324 MB**       | **6 GB**               | **64 MB**                 | **14.8 MB**                  |**6.4MB**                  |
-| **Reduction**         | **91% smaller**  | **97% smaller**        | **97% smaller**             | **95% smaller**              |**95% smaller**              |
-
-<!-- ### Memory Usage Comparison
-
-| System          j      | DPR(2M docs)     | RPJ-wiki(60M docs)    | Chat history()   |
-| --------------------- | ---------------- | ---------------- | ---------------- |
-| Traditional Vector DB(LLamaindex faiss) | x GB           | x GB            | x GB           |
-| **Leann**       | **xx MB** | **x GB** | **x GB** |
-| **Reduction**   | **x%**  | **x%**  | **x%**  |
-
-### Query Performance of LEANN
-
-| Backend             | Index Size | Query Time | Recall@3 |
-| ------------------- | ---------- | ---------- | --------- |
-| DiskANN             | 1M docs    | xms       | 0.95      |
-| HNSW                | 1M docs    | xms        | 0.95      | -->
-
-*Benchmarks run on Apple M3 Pro 36 GB*
-
+The evaluation script downloads data automatically on first run. The last three results were tested with partial personal data, and you can reproduce them with your own data!
 ## 🔬 Paper
 
 If you find Leann useful, please cite:
@@ -364,97 +477,25 @@ If you find Leann useful, please cite:
 
 ```bibtex
 @misc{wang2025leannlowstoragevectorindex,
-      title={LEANN: A Low-Storage Vector Index}, 
+      title={LEANN: A Low-Storage Vector Index},
       author={Yichuan Wang and Shu Liu and Zhifei Li and Yongji Wu and Ziming Mao and Yilong Zhao and Xiao Yan and Zhiying Xu and Yang Zhou and Ion Stoica and Sewon Min and Matei Zaharia and Joseph E. Gonzalez},
       year={2025},
       eprint={2506.08276},
       archivePrefix={arXiv},
       primaryClass={cs.DB},
-      url={https://arxiv.org/abs/2506.08276}, 
+      url={https://arxiv.org/abs/2506.08276},
 }
 ```
 
-## ✨ Features
+## ✨ [Detailed Features →](docs/features.md)
 
-### 🔥 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
-- **📈 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** - DiskANN, HNSW/FAISS with unified API
-
-### 🛠️ 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](test/build_mlx_index.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
-
-## 🤝 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
+## 🤝 [CONTRIBUTING →](docs/CONTRIBUTING.md)
 
 
-<!-- ## ❓ FAQ
-
-### Common Issues
-
-#### NCCL Topology Error
-
-**Problem**: You encounter `ncclTopoComputePaths` error during document processing:
-
-```
-ncclTopoComputePaths (system=<optimized out>, comm=comm@entry=0x5555a82fa3c0) at graph/paths.cc:688
-```
-
-**Solution**: Set these environment variables before running your script:
-
-```bash
-export NCCL_TOPO_DUMP_FILE=/tmp/nccl_topo.xml
-export NCCL_DEBUG=INFO
-export NCCL_DEBUG_SUBSYS=INIT,GRAPH
-export NCCL_IB_DISABLE=1
-export NCCL_NET_PLUGIN=none
-export NCCL_SOCKET_IFNAME=ens5
-``` -->
-
-## 📈 Roadmap
-
-### 🎯 Q2 2025
-
-- [X] DiskANN backend with MIPS/L2/Cosine support
-- [X] HNSW backend integration
-- [X] Real-time embedding pipeline
-- [X] Memory-efficient graph pruning
-
-### 🚀 Q3 2025
+## ❓ [FAQ →](docs/faq.md)
 
 
-- [ ] 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
+## 📈 [Roadmap →](docs/roadmap.md)
 
 ## 📄 License
 
@@ -462,11 +503,7 @@ MIT License - see [LICENSE](LICENSE) for details.
 
 ## 🙏 Acknowledgments
 
-- **Microsoft Research** for the DiskANN algorithm
-- **Meta AI** for FAISS and optimization insights
-- **HuggingFace** for the transformer ecosystem
-- **Our amazing contributors** who make this possible
-
+This work is done at [**Berkeley Sky Computing Lab**](https://sky.cs.berkeley.edu/)
 ---
 
 <p align="center">
@@ -476,4 +513,3 @@ MIT License - see [LICENSE](LICENSE) for details.
 <p align="center">
   Made with ❤️ by the Leann team
 </p>
-

demo.ipynb

@@ -1,35 +1,116 @@
 {
  "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# 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."
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
-    "from leann.api import LeannBuilder, LeannSearcher, LeannChat\n",
-    "# 1. Build index (no embeddings stored!)\n",
+    "# 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\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Build the index"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from leann.api import LeannBuilder\n",
+    "\n",
     "builder = LeannBuilder(backend_name=\"hnsw\")\n",
-    "builder.add_text(\"C# is a powerful programming language but it is not very popular\")\n",
-    "builder.add_text(\"Python is a powerful programming language and it is very popular\")\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\")\n",
-    "# 2. Search with real-time embeddings\n",
-    "searcher = LeannSearcher(\"knowledge.leann\")\n",
-    "results = searcher.search(\"programming languages\", top_k=2, recompute_beighbor_embeddings=True)\n",
-    "print(results)\n",
-    "\n",
-    "llm_config = {\"type\": \"ollama\", \"model\": \"qwen3:8b\"}\n",
-    "\n",
-    "chat = LeannChat(index_path=\"knowledge.leann\", llm_config=llm_config)\n",
-    "\n",
-    "response = chat.ask(\n",
-    "    \"Compare the two retrieved programming languages and say which one is more popular today. Respond in a single well-formed sentence.\",\n",
-    "    top_k=2,\n",
-    "    recompute_beighbor_embeddings=True,\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",
-    "print(response)"
+    "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(INDEX_PATH)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Search with real-time embeddings"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from leann.api import LeannSearcher\n",
+    "\n",
+    "searcher = LeannSearcher(INDEX_PATH)\n",
+    "results = searcher.search(\"programming languages\", top_k=2)\n",
+    "results"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Chat with LEANN using retrieved results"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from leann.api import LeannChat\n",
+    "\n",
+    "llm_config = {\n",
+    "    \"type\": \"hf\",\n",
+    "    \"model\": \"Qwen/Qwen3-0.6B\",\n",
+    "}\n",
+    "\n",
+    "chat = LeannChat(index_path=INDEX_PATH, llm_config=llm_config)\n",
+    "response = chat.ask(\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 +1,220 @@
+# 🤝 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** (already included when you run `uv sync`):
+   ```bash
+   uv pip install pre-commit
+   ```
+
+2. **Install the git hooks**:
+   ```bash
+   pre-commit install
+   ```
+
+3. **Run pre-commit manually** (optional):
+   ```bash
+   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
+# 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

@@ -0,0 +1,22 @@
+# Release Guide
+
+## Setup (One-time)
+
+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`
+
+## Release (One-click)
+
+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
+
+That's it! The workflow will automatically:
+- ✅ Update version in all packages
+- ✅ Build all packages
+- ✅ Publish to PyPI
+- ✅ Create GitHub tag and release
+
+Check progress: https://github.com/yichuan-w/LEANN/actions

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/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 +1,22 @@
+# ✨ 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
+- **📈 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** - DiskANN, HNSW/FAISS with unified API
+
+## 🛠️ 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](test/build_mlx_index.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/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 of [OpenAI embeddings with MIPS](../examples/main_cli_example.py).

docs/roadmap.md

@@ -0,0 +1,21 @@
+# 📈 Roadmap
+
+## 🎯 Q2 2025
+
+- [X] DiskANN backend with MIPS/L2/Cosine support
+- [X] HNSW backend integration
+- [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/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
@@ -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}
 
@@ -111,9 +110,8 @@ 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(
@@ -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:

examples/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.
-
-

examples/document_search.py

@@ -3,37 +3,47 @@
 Document search demo with recompute mode
 """
 
-import os
-from pathlib import Path
 import shutil
 import time
+from pathlib import Path
 
 # Import backend packages to trigger plugin registration
 try:
-    import leann_backend_diskann
-    import leann_backend_hnsw
+    import leann_backend_diskann  # noqa: F401
+    import leann_backend_hnsw  # noqa: F401
+
     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
+from leann.api import LeannBuilder, LeannChat, LeannSearcher
 
 
 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."},
+        {
+            "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"
@@ -44,94 +54,96 @@ def main():
 
     # --- 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
-    )
-    
+
+    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!")
+    print("\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}")
+        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[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"
+        "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...")
+
+    print("\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}")
-        
+            print(
+                f"  {i}. ID: {res.id}, Score: {res.score:.4f}, Text: '{res.text}', Metadata: {res.metadata}"
+            )
+
         # Compare results
-        print(f"\n--- Result comparison ---")
+        print("\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}")
-        
+            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)")
+            print("✅ Recompute mode working correctly (more accurate but slower)")
         else:
-            print(f"ℹ️  Recompute time is unusually fast, network recomputation may not be enabled")
-            
+            print("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...")
+    print("\n[PHASE 4] Starting chat session...")
     chat = LeannChat(index_path=INDEX_PATH)
     chat_response = chat.ask(query)
     print(f"You: {query}")
@@ -143,4 +155,4 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    main()

examples/email_data/LEANN_email_reader.py

@@ -1,11 +1,13 @@
-import os
 import email
+import os
 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
 
-def find_all_messages_directories(root: str = None) -> List[Path]:
+
+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.
@@ -14,111 +16,120 @@ def find_all_messages_directories(root: str = None) -> List[Path]:
         # 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):
+    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]:
+
+    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)
+        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:
+                        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)
+                        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')
-                                
+                                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:
+                                        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')
+                                            body += part.get_payload(decode=True).decode(
+                                                "utf-8", errors="ignore"
+                                            )
                                             # break
                                 else:
-                                    body = msg.get_payload(decode=True).decode('utf-8', errors='ignore')
-                                
+                                    body = msg.get_payload(decode=True).decode(
+                                        "utf-8", errors="ignore"
+                                    )
+
                                 # Create document content with metadata embedded in text
                                 doc_content = f"""
-[EMAIL METADATA]
-File: {filename}
-From: {from_addr}
-To: {to_addr}
-Subject: {subject}
-Date: {date}
-[END METADATA]
-
+[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 
+        return docs

examples/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

examples/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")
@@ -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")

examples/google_history_reader_leann.py

@@ -1,15 +1,17 @@
-import os
-import asyncio
 import argparse
+import asyncio
+import os
+
 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 leann.api import LeannBuilder, LeannChat
 from llama_index.core.node_parser import SentenceSplitter
 
 # dotenv.load_dotenv()  # handled above if python-dotenv is available
@@ -17,42 +19,51 @@ from llama_index.core.node_parser import SentenceSplitter
 # 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):
+
+def create_leann_index_from_multiple_chrome_profiles(
+    profile_dirs: list[Path],
+    index_path: str = "chrome_history_index.leann",
+    max_count: int = -1,
+    embedding_model: str = "facebook/contriever",
+    embedding_mode: str = "sentence-transformers",
+):
     """
     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
+        embedding_model: The embedding model to use
+        embedding_mode: The embedding backend mode
     """
     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 ---")
+        print("--- 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}")
-            
+            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
+                    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")
@@ -62,94 +73,111 @@ def create_leann_index_from_multiple_chrome_profiles(profile_dirs: List[Path], i
             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")
-        
+
+        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=25)
-        
+        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:
-                all_texts.append(node.get_content())
-        
+                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 ---")
+        print("--- 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...")
+        print("--- Building new LEANN index ---")
+
+        print("\n[PHASE 1] Building Leann index...")
 
         # Use HNSW backend for better macOS compatibility
+        # LeannBuilder will automatically detect normalized embeddings and set appropriate distance metric
         builder = LeannBuilder(
             backend_name="hnsw",
-            embedding_model="facebook/contriever",
-            graph_degree=32, 
+            embedding_model=embedding_model,
+            embedding_mode=embedding_mode,
+            graph_degree=32,
             complexity=64,
             is_compact=True,
             is_recompute=True,
-            num_threads=1  # Force single-threaded mode
+            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):
+
+def create_leann_index(
+    profile_path: str | None = None,
+    index_path: str = "chrome_history_index.leann",
+    max_count: int = 1000,
+    embedding_model: str = "facebook/contriever",
+    embedding_mode: str = "sentence-transformers",
+):
     """
     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
+        embedding_model: The embedding model to use
+        embedding_mode: The embedding backend mode
     """
     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 ---")
+        print("--- 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...")
+        print("--- Building new LEANN index ---")
+
+        print("\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
-        )
-        
+
+        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:
@@ -157,54 +185,57 @@ def create_leann_index(profile_path: str = None, index_path: str = "chrome_histo
             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 ---")
+        print("--- 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...")
+        print("--- Building new LEANN index ---")
+
+        print("\n[PHASE 1] Building Leann index...")
 
         # Use HNSW backend for better macOS compatibility
+        # LeannBuilder will automatically detect normalized embeddings and set appropriate distance metric
         builder = LeannBuilder(
             backend_name="hnsw",
-            embedding_model="facebook/contriever",
-            graph_degree=32, 
+            embedding_model=embedding_model,
+            embedding_mode=embedding_mode,
+            graph_degree=32,
             complexity=64,
             is_compact=True,
             is_recompute=True,
-            num_threads=1  # Force single-threaded mode
+            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...")
+    print("\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, 
+        query,
+        top_k=10,
         recompute_beighbor_embeddings=True,
         complexity=32,
         beam_width=1,
@@ -213,55 +244,104 @@ async def query_leann_index(index_path: str, query: str):
             "model": "gpt-4o",
             "api_key": os.getenv("OPENAI_API_KEY"),
         },
-        llm_kwargs={
-            "temperature": 0.0,
-            "max_tokens": 1000
-        }
+        llm_kwargs={"temperature": 0.0, "max_tokens": 1000},
     )
-    print(f"Leann: {chat_response}")
+
+    print(f"Leann chat response: \033[36m{chat_response}\033[0m")
+
 
 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="./chrome_history_index_leann_test",
-                       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)')
-    
+    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="./google_history_index",
+        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)",
+    )
+    parser.add_argument(
+        "--embedding-model",
+        type=str,
+        default="facebook/contriever",
+        help="The embedding model to use (e.g., 'facebook/contriever', 'text-embedding-3-small')",
+    )
+    parser.add_argument(
+        "--embedding-mode",
+        type=str,
+        default="sentence-transformers",
+        choices=["sentence-transformers", "openai", "mlx"],
+        help="The embedding backend mode",
+    )
+    parser.add_argument(
+        "--use-existing-index",
+        action="store_true",
+        help="Use existing index without rebuilding",
+    )
+
     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.")
+
+    if args.use_existing_index:
+        # Use existing index without rebuilding
+        if not Path(INDEX_PATH).exists():
+            print(f"Error: Index file not found at {INDEX_PATH}")
             return
+        print(f"Using existing index at {INDEX_PATH}")
+        index_path = INDEX_PATH
     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)
-    
+        # 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, args.embedding_model, args.embedding_mode
+        )
+
     if index_path:
         if args.query:
             # Run single query
@@ -270,12 +350,13 @@ async def main():
             # Example queries
             queries = [
                 "What websites did I visit about machine learning?",
-                "Find my search history about programming"
+                "Find my search history about programming",
             ]
-            
+
             for query in queries:
-                print("\n" + "="*60)
+                print("\n" + "=" * 60)
                 await query_leann_index(index_path, query)
 
+
 if __name__ == "__main__":
-    asyncio.run(main()) 
+    asyncio.run(main())

examples/history_data/__init__.py

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

examples/history_data/history.py

@@ -1,122 +1,121 @@
-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
-                
+
                 # Create document content with metadata embedded in text
                 doc_content = f"""
-[BROWSING HISTORY METADATA]
-URL: {url}
-Title: {title}
-Last Visit: {last_visit}
-Visit Count: {visit_count}
-Typed Count: {typed_count}
-Hidden: {hidden}
-[END METADATA]
-
-Title: {title}
-URL: {url}
-Last visited: {last_visit}
+[Title]: {title}
+[URL of the page]: {url}
+[Last visited time]: {last_visit}
+[Visit times]: {visit_count}
+[Typed times]: {typed_count}
 """
-                
+
                 # Create document with embedded metadata
-                doc = Document(text=doc_content, metadata={})
+                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}")
             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":
@@ -124,53 +123,59 @@ Last visited: {last_visit}
                 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}")

examples/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:
@@ -69,303 +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)
-                    time_str = timestamp.strftime('%H:%M:%S')
-                except:
+                    # change to YYYY-MM-DD HH:MM:SS
+                    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}")
-        
+
+            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!
         doc_content = f"""
-Contact: {contact_name}
-
 {concatenated_text}
 """
-        return doc_content
-    
-    def load_data(self, input_dir: str = None, **load_kwargs: Any) -> List[Document]:
+        return doc_content, contact_name
+
+    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:
@@ -377,97 +406,104 @@ Contact: {contact_name}
                 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)
+        load_kwargs.get("max_length", 1000)
+        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, 
+                            readable_messages,
+                            max_length=-1,
                             time_window_minutes=-1,
-                            overlap_messages=0  # Keep 2 messages overlap between groups
+                            overlap_messages=0,  # Keep 2 messages 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 = self._create_concatenated_content(message_group, contact_name)
-                            doc = Document(text=doc_content, metadata={})
+
+                            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:
@@ -476,17 +512,17 @@ Contact: {contact_name}
                                 # 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}
@@ -494,57 +530,64 @@ 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={})
                             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_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
@@ -553,36 +596,36 @@ 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()
@@ -592,83 +635,90 @@ 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}")
@@ -679,18 +729,18 @@ 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"),
@@ -698,23 +748,25 @@ 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

examples/mail_reader_leann.py

@@ -1,33 +1,42 @@
+import argparse
+import asyncio
 import os
 import sys
-import asyncio
-import dotenv
-import argparse
 from pathlib import Path
-from typing import List, Any
+
+import dotenv
 
 # Add the project root to Python path so we can import from examples
 project_root = Path(__file__).parent.parent
 sys.path.insert(0, str(project_root))
 
-from leann.api import LeannBuilder, LeannSearcher, LeannChat
+from leann.api import LeannBuilder, LeannChat
 from llama_index.core.node_parser import SentenceSplitter
 
 dotenv.load_dotenv()
 
+
 # Auto-detect user's mail path
 def get_mail_path():
     """Get the mail path for the current user"""
     home_dir = os.path.expanduser("~")
     return os.path.join(home_dir, "Library", "Mail")
 
-# Default mail path for macOS
-# DEFAULT_MAIL_PATH = "/Users/yichuan/Library/Mail/V10/0FCA0879-FD8C-4B7E-83BF-FDDA930791C5/[Gmail].mbox/All Mail.mbox/78BA5BE1-8819-4F9A-9613-EB63772F1DD0/Data"
 
-def create_leann_index_from_multiple_sources(messages_dirs: List[Path], index_path: str = "mail_index.leann", max_count: int = -1, include_html: bool = False, embedding_model: str = "facebook/contriever"):
+# Default mail path for macOS
+DEFAULT_MAIL_PATH = "/Users/yichuan/Library/Mail/V10/0FCA0879-FD8C-4B7E-83BF-FDDA930791C5/[Gmail].mbox/All Mail.mbox/78BA5BE1-8819-4F9A-9613-EB63772F1DD0/Data"
+
+
+def create_leann_index_from_multiple_sources(
+    messages_dirs: list[Path],
+    index_path: str = "mail_index.leann",
+    max_count: int = -1,
+    include_html: bool = False,
+    embedding_model: str = "facebook/contriever",
+):
     """
     Create LEANN index from multiple mail data sources.
-    
+
     Args:
         messages_dirs: List of Path objects pointing to Messages directories
         index_path: Path to save the LEANN index
@@ -35,31 +44,32 @@ def create_leann_index_from_multiple_sources(messages_dirs: List[Path], index_pa
         include_html: Whether to include HTML content in email processing
     """
     print("Creating LEANN index from multiple mail data sources...")
-    
+
     # Load documents using EmlxReader from LEANN_email_reader
     from examples.email_data.LEANN_email_reader import EmlxReader
+
     reader = EmlxReader(include_html=include_html)
     # from email_data.email import EmlxMboxReader
     # from pathlib import Path
     # reader = EmlxMboxReader()
     INDEX_DIR = Path(index_path).parent
-    
+
     if not INDEX_DIR.exists():
-        print(f"--- Index directory not found, building new index ---")
+        print("--- Index directory not found, building new index ---")
         all_documents = []
         total_processed = 0
-        
+
         # Process each Messages directory
         for i, messages_dir in enumerate(messages_dirs):
-            print(f"\nProcessing Messages directory {i+1}/{len(messages_dirs)}: {messages_dir}")
-            
+            print(f"\nProcessing Messages directory {i + 1}/{len(messages_dirs)}: {messages_dir}")
+
             try:
                 documents = reader.load_data(messages_dir)
                 if documents:
                     print(f"Loaded {len(documents)} email documents from {messages_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")
@@ -69,61 +79,74 @@ def create_leann_index_from_multiple_sources(messages_dirs: List[Path], index_pa
             except Exception as e:
                 print(f"Error processing {messages_dir}: {e}")
                 continue
-        
+
         if not all_documents:
             print("No documents loaded from any source. Exiting.")
             return None
-        
-        print(f"\nTotal loaded {len(all_documents)} email documents from {len(messages_dirs)} directories")
-        
+
+        print(
+            f"\nTotal loaded {len(all_documents)} email documents from {len(messages_dirs)} directories and starting to split them into chunks"
+        )
+
         # 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 all_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(all_documents)} documents")
-        
+                text = node.get_content()
+                # text = '[subject] ' + doc.metadata["subject"] + '\n' + text
+                all_texts.append(text)
+
+        print(
+            f"Finished splitting {len(all_documents)} documents into {len(all_texts)} text chunks"
+        )
+
         # Create LEANN index directory
 
-        print(f"--- Index directory not found, building new index ---")
+        print("--- 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...")
+        print("--- Building new LEANN index ---")
+
+        print("\n[PHASE 1] Building Leann index...")
 
         # Use HNSW backend for better macOS compatibility
         builder = LeannBuilder(
             backend_name="hnsw",
             embedding_model=embedding_model,
-            graph_degree=32, 
+            graph_degree=32,
             complexity=64,
             is_compact=True,
             is_recompute=True,
-            num_threads=1  # Force single-threaded mode
+            num_threads=1,  # Force single-threaded mode
         )
 
         print(f"Adding {len(all_texts)} email 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(mail_path: str, index_path: str = "mail_index.leann", max_count: int = 1000, include_html: bool = False, embedding_model: str = "facebook/contriever"):
+
+def create_leann_index(
+    mail_path: str,
+    index_path: str = "mail_index.leann",
+    max_count: int = 1000,
+    include_html: bool = False,
+    embedding_model: str = "facebook/contriever",
+):
     """
     Create LEANN index from mail data.
-    
+
     Args:
         mail_path: Path to the mail directory
         index_path: Path to save the LEANN index
@@ -132,32 +155,33 @@ def create_leann_index(mail_path: str, index_path: str = "mail_index.leann", max
     """
     print("Creating LEANN index from mail data...")
     INDEX_DIR = Path(index_path).parent
-    
+
     if not INDEX_DIR.exists():
-        print(f"--- Index directory not found, building new index ---")
+        print("--- 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...")
+        print("--- Building new LEANN index ---")
+
+        print("\n[PHASE 1] Building Leann index...")
 
         # Load documents using EmlxReader from LEANN_email_reader
         from examples.email_data.LEANN_email_reader import EmlxReader
+
         reader = EmlxReader(include_html=include_html)
         # from email_data.email import EmlxMboxReader
         # from pathlib import Path
         # reader = EmlxMboxReader()
         documents = reader.load_data(Path(mail_path))
-        
+
         if not documents:
             print("No documents loaded. Exiting.")
             return None
-        
+
         print(f"Loaded {len(documents)} email documents")
-        
+
         # Create text splitter with 256 chunk size
-        text_splitter = SentenceSplitter(chunk_size=256, chunk_overlap=25)
-        
+        text_splitter = SentenceSplitter(chunk_size=256, chunk_overlap=128)
+
         # Convert Documents to text strings and chunk them
         all_texts = []
         for doc in documents:
@@ -165,108 +189,139 @@ def create_leann_index(mail_path: str, index_path: str = "mail_index.leann", max
             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 ---")
+        print("--- 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...")
+        print("--- Building new LEANN index ---")
+
+        print("\n[PHASE 1] Building Leann index...")
 
         # Use HNSW backend for better macOS compatibility
         builder = LeannBuilder(
             backend_name="hnsw",
             embedding_model=embedding_model,
-            graph_degree=32, 
+            graph_degree=32,
             complexity=64,
             is_compact=True,
             is_recompute=True,
-            num_threads=1  # Force single-threaded mode
+            num_threads=1,  # Force single-threaded mode
         )
 
         print(f"Adding {len(all_texts)} email 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,
-                     llm_config={"type": "openai", "model": "gpt-4o"})
-    
+    print("\n[PHASE 2] Starting Leann chat session...")
+    chat = LeannChat(index_path=index_path, llm_config={"type": "openai", "model": "gpt-4o"})
+
     print(f"You: {query}")
     import time
-    start_time = time.time()
+
+    time.time()
     chat_response = chat.ask(
-        query, 
-        top_k=10, 
+        query,
+        top_k=20,
         recompute_beighbor_embeddings=True,
-        complexity=12,
+        complexity=32,
         beam_width=1,
-        
     )
-    end_time = time.time()
-    print(f"Time taken: {end_time - start_time} seconds")
-    print(f"Leann: {chat_response}")
+    time.time()
+    # print(f"Time taken: {end_time - start_time} seconds")
+    # highlight the answer
+    print(f"Leann chat response: \033[36m{chat_response}\033[0m")
+
 
 async def main():
     # Parse command line arguments
-    parser = argparse.ArgumentParser(description='LEANN Mail Reader - Create and query email index')
+    parser = argparse.ArgumentParser(description="LEANN Mail Reader - Create and query email index")
     # Remove --mail-path argument and auto-detect all Messages directories
     # Remove DEFAULT_MAIL_PATH
-    parser.add_argument('--index-dir', type=str, default="./mail_index_leann_raw_text_all_dicts",
-                       help='Directory to store the LEANN index (default: ./mail_index_leann_raw_text_all_dicts)')
-    parser.add_argument('--max-emails', type=int, default=1000,
-                       help='Maximum number of emails to process (-1 means all)')
-    parser.add_argument('--query', type=str, default="Give me some funny advertisement about apple or other companies",
-                       help='Single query to run (default: runs example queries)')
-    parser.add_argument('--include-html', action='store_true', default=False,
-                       help='Include HTML content in email processing (default: False)')
-    parser.add_argument('--embedding-model', type=str, default="facebook/contriever",
-                       help='Embedding model to use (default: facebook/contriever)')
-    
+    parser.add_argument(
+        "--index-dir",
+        type=str,
+        default="./mail_index",
+        help="Directory to store the LEANN index (default: ./mail_index_leann_raw_text_all_dicts)",
+    )
+    parser.add_argument(
+        "--max-emails",
+        type=int,
+        default=1000,
+        help="Maximum number of emails to process (-1 means all)",
+    )
+    parser.add_argument(
+        "--query",
+        type=str,
+        default="Give me some funny advertisement about apple or other companies",
+        help="Single query to run (default: runs example queries)",
+    )
+    parser.add_argument(
+        "--include-html",
+        action="store_true",
+        default=False,
+        help="Include HTML content in email processing (default: False)",
+    )
+    parser.add_argument(
+        "--embedding-model",
+        type=str,
+        default="facebook/contriever",
+        help="Embedding model to use (default: facebook/contriever)",
+    )
+
     args = parser.parse_args()
 
     print(f"args: {args}")
-    
+
     # Automatically find all Messages directories under the current user's Mail directory
     from examples.email_data.LEANN_email_reader import find_all_messages_directories
+
     mail_path = get_mail_path()
     print(f"Searching for email data in: {mail_path}")
     messages_dirs = find_all_messages_directories(mail_path)
-    
-    print('len(messages_dirs): ', len(messages_dirs))
-    
-    
+    # messages_dirs = find_all_messages_directories(DEFAULT_MAIL_PATH)
+    # messages_dirs = [DEFAULT_MAIL_PATH]
+    # messages_dirs = messages_dirs[:1]
+
+    print("len(messages_dirs): ", len(messages_dirs))
+
     if not messages_dirs:
         print("No Messages directories found. Exiting.")
         return
-    
+
     INDEX_DIR = Path(args.index_dir)
     INDEX_PATH = str(INDEX_DIR / "mail_documents.leann")
     print(f"Index directory: {INDEX_DIR}")
     print(f"Found {len(messages_dirs)} Messages directories.")
-    
+
     # Create or load the LEANN index from all sources
-    index_path = create_leann_index_from_multiple_sources(messages_dirs, INDEX_PATH, args.max_emails, args.include_html, args.embedding_model)
-    
+    index_path = create_leann_index_from_multiple_sources(
+        messages_dirs,
+        INDEX_PATH,
+        args.max_emails,
+        args.include_html,
+        args.embedding_model,
+    )
+
     if index_path:
         if args.query:
             # Run single query
@@ -276,11 +331,12 @@ async def main():
             queries = [
                 "Hows Berkeley Graduate Student Instructor",
                 "how's the icloud related advertisement saying",
-                "Whats the number of class recommend to take per semester for incoming EECS students"
+                "Whats the number of class recommend to take per semester for incoming EECS students",
             ]
             for query in queries:
-                print("\n" + "="*60)
+                print("\n" + "=" * 60)
                 await query_leann_index(index_path, query)
 
+
 if __name__ == "__main__":
-    asyncio.run(main()) 
+    asyncio.run(main())

examples/mail_reader_llamaindex.py

@@ -1,26 +1,30 @@
+import argparse
 import os
 import sys
-import argparse
 from pathlib import Path
-from typing import List, Any
 
 # Add the project root to Python path so we can import from examples
 project_root = Path(__file__).parent.parent
 sys.path.insert(0, str(project_root))
 
-from llama_index.core import VectorStoreIndex, StorageContext
+import torch
+from llama_index.core import StorageContext, VectorStoreIndex
 from llama_index.core.node_parser import SentenceSplitter
 
 # --- EMBEDDING MODEL ---
 from llama_index.embeddings.huggingface import HuggingFaceEmbedding
-import torch
 
 # --- END EMBEDDING MODEL ---
-
 # Import EmlxReader from the new module
 from examples.email_data.LEANN_email_reader import EmlxReader
 
-def create_and_save_index(mail_path: str, save_dir: str = "mail_index_embedded", max_count: int = 1000, include_html: bool = False):
+
+def create_and_save_index(
+    mail_path: str,
+    save_dir: str = "mail_index_embedded",
+    max_count: int = 1000,
+    include_html: bool = False,
+):
     print("Creating index from mail data with embedded metadata...")
     documents = EmlxReader(include_html=include_html).load_data(mail_path, max_count=max_count)
     if not documents:
@@ -30,7 +34,7 @@ def create_and_save_index(mail_path: str, save_dir: str = "mail_index_embedded",
     # Use facebook/contriever as the embedder
     embed_model = HuggingFaceEmbedding(model_name="facebook/contriever")
     # set on device
-    import torch
+
     if torch.cuda.is_available():
         embed_model._model.to("cuda")
     # set mps
@@ -39,21 +43,19 @@ def create_and_save_index(mail_path: str, save_dir: str = "mail_index_embedded",
     else:
         embed_model._model.to("cpu")
     index = VectorStoreIndex.from_documents(
-        documents,
-        transformations=[text_splitter],
-        embed_model=embed_model
+        documents, transformations=[text_splitter], embed_model=embed_model
     )
     os.makedirs(save_dir, exist_ok=True)
     index.storage_context.persist(persist_dir=save_dir)
     print(f"Index saved to {save_dir}")
     return index
 
+
 def load_index(save_dir: str = "mail_index_embedded"):
     try:
         storage_context = StorageContext.from_defaults(persist_dir=save_dir)
         index = VectorStoreIndex.from_vector_store(
-            storage_context.vector_store,
-            storage_context=storage_context
+            storage_context.vector_store, storage_context=storage_context
         )
         print(f"Index loaded from {save_dir}")
         return index
@@ -61,6 +63,7 @@ def load_index(save_dir: str = "mail_index_embedded"):
         print(f"Error loading index: {e}")
         return None
 
+
 def query_index(index, query: str):
     if index is None:
         print("No index available for querying.")
@@ -70,39 +73,63 @@ def query_index(index, query: str):
     print(f"Query: {query}")
     print(f"Response: {response}")
 
+
 def main():
     # Parse command line arguments
-    parser = argparse.ArgumentParser(description='LlamaIndex Mail Reader - Create and query email index')
-    parser.add_argument('--mail-path', type=str, 
-                       default="/Users/yichuan/Library/Mail/V10/0FCA0879-FD8C-4B7E-83BF-FDDA930791C5/[Gmail].mbox/All Mail.mbox/78BA5BE1-8819-4F9A-9613-EB63772F1DD0/Data/9/Messages",
-                       help='Path to mail data directory')
-    parser.add_argument('--save-dir', type=str, default="mail_index_embedded",
-                       help='Directory to store the index (default: mail_index_embedded)')
-    parser.add_argument('--max-emails', type=int, default=10000,
-                       help='Maximum number of emails to process')
-    parser.add_argument('--include-html', action='store_true', default=False,
-                       help='Include HTML content in email processing (default: False)')
-    
+    parser = argparse.ArgumentParser(
+        description="LlamaIndex Mail Reader - Create and query email index"
+    )
+    parser.add_argument(
+        "--mail-path",
+        type=str,
+        default="/Users/yichuan/Library/Mail/V10/0FCA0879-FD8C-4B7E-83BF-FDDA930791C5/[Gmail].mbox/All Mail.mbox/78BA5BE1-8819-4F9A-9613-EB63772F1DD0/Data/9/Messages",
+        help="Path to mail data directory",
+    )
+    parser.add_argument(
+        "--save-dir",
+        type=str,
+        default="mail_index_embedded",
+        help="Directory to store the index (default: mail_index_embedded)",
+    )
+    parser.add_argument(
+        "--max-emails",
+        type=int,
+        default=10000,
+        help="Maximum number of emails to process",
+    )
+    parser.add_argument(
+        "--include-html",
+        action="store_true",
+        default=False,
+        help="Include HTML content in email processing (default: False)",
+    )
+
     args = parser.parse_args()
-    
+
     mail_path = args.mail_path
     save_dir = args.save_dir
-    
+
     if os.path.exists(save_dir) and os.path.exists(os.path.join(save_dir, "vector_store.json")):
         print("Loading existing index...")
         index = load_index(save_dir)
     else:
         print("Creating new index...")
-        index = create_and_save_index(mail_path, save_dir, max_count=args.max_emails, include_html=args.include_html)
+        index = create_and_save_index(
+            mail_path,
+            save_dir,
+            max_count=args.max_emails,
+            include_html=args.include_html,
+        )
     if index:
         queries = [
             "Hows Berkeley Graduate Student Instructor",
             "how's the icloud related advertisement saying",
-            "Whats the number of class recommend to take per semester for incoming EECS students"
+            "Whats the number of class recommend to take per semester for incoming EECS students",
         ]
         for query in queries:
-            print("\n" + "="*50)
+            print("\n" + "=" * 50)
             query_index(index, query)
 
+
 if __name__ == "__main__":
-    main() 
+    main()

examples/main_cli_example.py

@@ -1,30 +1,13 @@
 import argparse
-from llama_index.core import SimpleDirectoryReader, Settings
-from llama_index.core.node_parser import SentenceSplitter
 import asyncio
-import dotenv
-from leann.api import LeannBuilder, LeannSearcher, LeannChat
-import shutil
 from pathlib import Path
 
-dotenv.load_dotenv()
+import dotenv
+from leann.api import LeannBuilder, LeannChat
+from llama_index.core import SimpleDirectoryReader
+from llama_index.core.node_parser import SentenceSplitter
 
-node_parser = SentenceSplitter(
-    chunk_size=256, chunk_overlap=128, separator=" ", paragraph_separator="\n\n"
-)
-print("Loading documents...")
-documents = SimpleDirectoryReader(
-    "examples/data",
-    recursive=True,
-    encoding="utf-8",
-    required_exts=[".pdf", ".txt", ".md"],
-).load_data(show_progress=True)
-print("Documents loaded.")
-all_texts = []
-for doc in documents:
-    nodes = node_parser.get_nodes_from_documents([doc])
-    for node in nodes:
-        all_texts.append(node.get_content())
+dotenv.load_dotenv()
 
 
 async def main(args):
@@ -32,14 +15,37 @@ async def main(args):
     INDEX_PATH = str(INDEX_DIR / "pdf_documents.leann")
 
     if not INDEX_DIR.exists():
-        print(f"--- Index directory not found, building new index ---")
+        node_parser = SentenceSplitter(
+            chunk_size=256, chunk_overlap=128, separator=" ", paragraph_separator="\n\n"
+        )
 
-        print(f"\n[PHASE 1] Building Leann index...")
+        print("Loading documents...")
+        documents = SimpleDirectoryReader(
+            args.data_dir,
+            recursive=True,
+            encoding="utf-8",
+            required_exts=[".pdf", ".txt", ".md"],
+        ).load_data(show_progress=True)
+        print("Documents loaded.")
+        all_texts = []
+        for doc in documents:
+            nodes = node_parser.get_nodes_from_documents([doc])
+            if nodes:
+                all_texts.extend(node.get_content() for node in nodes)
+
+        print("--- Index directory not found, building new index ---")
+
+        print("\n[PHASE 1] Building Leann index...")
+
+        # LeannBuilder now automatically detects normalized embeddings and sets appropriate distance metric
+        print(f"Using {args.embedding_model} with {args.embedding_mode} mode")
 
         # Use HNSW backend for better macOS compatibility
         builder = LeannBuilder(
             backend_name="hnsw",
-            embedding_model="facebook/contriever",
+            embedding_model=args.embedding_model,
+            embedding_mode=args.embedding_mode,
+            # distance_metric is automatically set based on embedding model
             graph_degree=32,
             complexity=64,
             is_compact=True,
@@ -56,43 +62,61 @@ async def main(args):
     else:
         print(f"--- Using existing index at {INDEX_DIR} ---")
 
-    print(f"\n[PHASE 2] Starting Leann chat session...")
+    print("\n[PHASE 2] Starting Leann chat session...")
 
-    # llm_config = {"type": "hf", "model": "Qwen/Qwen3-4B"}
-    llm_config = {"type": "ollama", "model": "qwen3:8b"}
+    # Build llm_config based on command line arguments
+    if args.llm == "simulated":
+        llm_config = {"type": "simulated"}
+    elif args.llm == "ollama":
+        llm_config = {"type": "ollama", "model": args.model, "host": args.host}
+    elif args.llm == "hf":
+        llm_config = {"type": "hf", "model": args.model}
+    elif args.llm == "openai":
+        llm_config = {"type": "openai", "model": args.model}
+    else:
+        raise ValueError(f"Unknown LLM type: {args.llm}")
+
+    print(f"Using LLM: {args.llm} with model: {args.model if args.llm != 'simulated' else 'N/A'}")
 
     chat = LeannChat(index_path=INDEX_PATH, llm_config=llm_config)
-
-    query = "Based on the paper, what are the main techniques LEANN explores to reduce the storage overhead and DLPM explore to achieve Fairness and Efiiciency trade-off?"
-
     # query = (
-    #     "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"
     # )
+    query = args.query
 
     print(f"You: {query}")
-    chat_response = chat.ask(
-        query, top_k=20, recompute_beighbor_embeddings=True, complexity=32
-    )
-    print(f"Leann: {chat_response}")
+    chat_response = chat.ask(query, top_k=20, recompute_embeddings=True, complexity=32)
+    print(f"Leann chat response: \033[36m{chat_response}\033[0m")
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Run Leann Chat with various LLM backends."
-    )
+    parser = argparse.ArgumentParser(description="Run Leann Chat with various LLM backends.")
     parser.add_argument(
         "--llm",
         type=str,
-        default="hf",
+        default="openai",
         choices=["simulated", "ollama", "hf", "openai"],
         help="The LLM backend to use.",
     )
     parser.add_argument(
         "--model",
         type=str,
-        default="Qwen/Qwen3-0.6B",
+        default="gpt-4o",
         help="The model name to use (e.g., 'llama3:8b' for ollama, 'deepseek-ai/deepseek-llm-7b-chat' for hf, 'gpt-4o' for openai).",
     )
+    parser.add_argument(
+        "--embedding-model",
+        type=str,
+        default="facebook/contriever",
+        help="The embedding model to use (e.g., 'facebook/contriever', 'text-embedding-3-small').",
+    )
+    parser.add_argument(
+        "--embedding-mode",
+        type=str,
+        default="sentence-transformers",
+        choices=["sentence-transformers", "openai", "mlx"],
+        help="The embedding backend mode.",
+    )
     parser.add_argument(
         "--host",
         type=str,
@@ -105,6 +129,18 @@ if __name__ == "__main__":
         default="./test_doc_files",
         help="Directory where the Leann index will be stored.",
     )
+    parser.add_argument(
+        "--data-dir",
+        type=str,
+        default="examples/data",
+        help="Directory containing documents to index (PDF, TXT, MD files).",
+    )
+    parser.add_argument(
+        "--query",
+        type=str,
+        default="Based on the paper, what are the main techniques LEANN explores to reduce the storage overhead and DLPM explore to achieve Fairness and Efiiciency trade-off?",
+        help="The query to ask the Leann chat system.",
+    )
     args = parser.parse_args()
 
     asyncio.run(main(args))

examples/multi_vector_aggregator.py

@@ -14,48 +14,55 @@ Key features:
 - Document-level result consolidation
 """
 
-import numpy as np
-from typing import List, Dict, Any, Tuple, Optional
-from dataclasses import dataclass
 from collections import defaultdict
-import json
+from dataclasses import dataclass
+from typing import Any
+
+import numpy as np
+
 
 @dataclass
 class PatchResult:
     """Represents a single patch search result."""
+
     patch_id: int
     image_name: str
     image_path: str
-    coordinates: Tuple[int, int, int, int]  # (x1, y1, x2, y2)
+    coordinates: tuple[int, int, int, int]  # (x1, y1, x2, y2)
     score: float
     attention_score: float
     scale: float
-    metadata: Dict[str, Any]
+    metadata: dict[str, Any]
+
 
 @dataclass
 class AggregatedResult:
     """Represents an aggregated document-level result."""
+
     image_name: str
     image_path: str
     doc_score: float
     patch_count: int
     best_patch: PatchResult
-    all_patches: List[PatchResult]
+    all_patches: list[PatchResult]
     aggregation_method: str
-    spatial_clusters: Optional[List[List[PatchResult]]] = None
+    spatial_clusters: list[list[PatchResult]] | None = None
+
 
 class MultiVectorAggregator:
     """
     Aggregates multiple patch-level results into document-level results.
     """
-    
-    def __init__(self, 
-                 aggregation_method: str = "maxsim",
-                 spatial_clustering: bool = True,
-                 cluster_distance_threshold: float = 100.0):
+
+    def __init__(
+        self,
+        aggregation_method: str = "maxsim",
+        spatial_clustering: bool = True,
+        cluster_distance_threshold: float = 100.0,
+    ):
         """
         Initialize the aggregator.
-        
+
         Args:
             aggregation_method: "maxsim", "voting", "weighted", or "mean"
             spatial_clustering: Whether to cluster spatially close patches
@@ -64,23 +71,23 @@ class MultiVectorAggregator:
         self.aggregation_method = aggregation_method
         self.spatial_clustering = spatial_clustering
         self.cluster_distance_threshold = cluster_distance_threshold
-    
-    def aggregate_results(self, 
-                         search_results: List[Dict[str, Any]], 
-                         top_k: int = 10) -> List[AggregatedResult]:
+
+    def aggregate_results(
+        self, search_results: list[dict[str, Any]], top_k: int = 10
+    ) -> list[AggregatedResult]:
         """
         Aggregate patch-level search results into document-level results.
-        
+
         Args:
             search_results: List of search results from LeannSearcher
             top_k: Number of top documents to return
-            
+
         Returns:
             List of aggregated document results
         """
         # Group results by image
         image_groups = defaultdict(list)
-        
+
         for result in search_results:
             metadata = result.metadata
             if "image_name" in metadata and "patch_id" in metadata:
@@ -92,55 +99,57 @@ class MultiVectorAggregator:
                     score=result.score,
                     attention_score=metadata.get("attention_score", 0.0),
                     scale=metadata.get("scale", 1.0),
-                    metadata=metadata
+                    metadata=metadata,
                 )
                 image_groups[metadata["image_name"]].append(patch_result)
-        
+
         # Aggregate each image group
         aggregated_results = []
         for image_name, patches in image_groups.items():
             if len(patches) == 0:
                 continue
-                
+
             agg_result = self._aggregate_image_patches(image_name, patches)
             aggregated_results.append(agg_result)
-        
+
         # Sort by aggregated score and return top-k
         aggregated_results.sort(key=lambda x: x.doc_score, reverse=True)
         return aggregated_results[:top_k]
-    
-    def _aggregate_image_patches(self, image_name: str, patches: List[PatchResult]) -> AggregatedResult:
+
+    def _aggregate_image_patches(
+        self, image_name: str, patches: list[PatchResult]
+    ) -> AggregatedResult:
         """Aggregate patches for a single image."""
-        
+
         if self.aggregation_method == "maxsim":
             doc_score = max(patch.score for patch in patches)
             best_patch = max(patches, key=lambda p: p.score)
-            
+
         elif self.aggregation_method == "voting":
             # Count patches above threshold
             threshold = np.percentile([p.score for p in patches], 75)
             doc_score = sum(1 for patch in patches if patch.score >= threshold)
             best_patch = max(patches, key=lambda p: p.score)
-            
+
         elif self.aggregation_method == "weighted":
             # Weight by attention scores
             total_weighted_score = sum(p.score * p.attention_score for p in patches)
             total_weights = sum(p.attention_score for p in patches)
             doc_score = total_weighted_score / max(total_weights, 1e-8)
             best_patch = max(patches, key=lambda p: p.score * p.attention_score)
-            
+
         elif self.aggregation_method == "mean":
             doc_score = np.mean([patch.score for patch in patches])
             best_patch = max(patches, key=lambda p: p.score)
-            
+
         else:
             raise ValueError(f"Unknown aggregation method: {self.aggregation_method}")
-        
+
         # Spatial clustering if enabled
         spatial_clusters = None
         if self.spatial_clustering:
             spatial_clusters = self._cluster_patches_spatially(patches)
-        
+
         return AggregatedResult(
             image_name=image_name,
             image_path=patches[0].image_path,
@@ -149,23 +158,23 @@ class MultiVectorAggregator:
             best_patch=best_patch,
             all_patches=sorted(patches, key=lambda p: p.score, reverse=True),
             aggregation_method=self.aggregation_method,
-            spatial_clusters=spatial_clusters
+            spatial_clusters=spatial_clusters,
         )
-    
-    def _cluster_patches_spatially(self, patches: List[PatchResult]) -> List[List[PatchResult]]:
+
+    def _cluster_patches_spatially(self, patches: list[PatchResult]) -> list[list[PatchResult]]:
         """Cluster patches that are spatially close to each other."""
         if len(patches) <= 1:
             return [patches]
-        
+
         clusters = []
         remaining_patches = patches.copy()
-        
+
         while remaining_patches:
             # Start new cluster with highest scoring remaining patch
             seed_patch = max(remaining_patches, key=lambda p: p.score)
             current_cluster = [seed_patch]
             remaining_patches.remove(seed_patch)
-            
+
             # Add nearby patches to cluster
             added_to_cluster = True
             while added_to_cluster:
@@ -175,145 +184,177 @@ class MultiVectorAggregator:
                         current_cluster.append(patch)
                         remaining_patches.remove(patch)
                         added_to_cluster = True
-            
+
             clusters.append(current_cluster)
-        
+
         return sorted(clusters, key=lambda cluster: max(p.score for p in cluster), reverse=True)
-    
-    def _is_patch_nearby(self, patch: PatchResult, cluster: List[PatchResult]) -> bool:
+
+    def _is_patch_nearby(self, patch: PatchResult, cluster: list[PatchResult]) -> bool:
         """Check if a patch is spatially close to any patch in the cluster."""
         patch_center = self._get_patch_center(patch.coordinates)
-        
+
         for cluster_patch in cluster:
             cluster_center = self._get_patch_center(cluster_patch.coordinates)
-            distance = np.sqrt((patch_center[0] - cluster_center[0])**2 + 
-                             (patch_center[1] - cluster_center[1])**2)
-            
+            distance = np.sqrt(
+                (patch_center[0] - cluster_center[0]) ** 2
+                + (patch_center[1] - cluster_center[1]) ** 2
+            )
+
             if distance <= self.cluster_distance_threshold:
                 return True
-        
+
         return False
-    
-    def _get_patch_center(self, coordinates: Tuple[int, int, int, int]) -> Tuple[float, float]:
+
+    def _get_patch_center(self, coordinates: tuple[int, int, int, int]) -> tuple[float, float]:
         """Get center point of a patch."""
         x1, y1, x2, y2 = coordinates
         return ((x1 + x2) / 2, (y1 + y2) / 2)
-    
-    def print_aggregated_results(self, results: List[AggregatedResult], max_patches_per_doc: int = 3):
+
+    def print_aggregated_results(
+        self, results: list[AggregatedResult], max_patches_per_doc: int = 3
+    ):
         """Pretty print aggregated results."""
         print(f"\n🔍 Aggregated Results (method: {self.aggregation_method})")
         print("=" * 80)
-        
+
         for i, result in enumerate(results):
-            print(f"\n{i+1}. {result.image_name}")
+            print(f"\n{i + 1}. {result.image_name}")
             print(f"   Doc Score: {result.doc_score:.4f} | Patches: {result.patch_count}")
             print(f"   Path: {result.image_path}")
-            
+
             # Show best patch
             best = result.best_patch
-            print(f"   🌟 Best Patch: #{best.patch_id} at {best.coordinates} (score: {best.score:.4f})")
-            
+            print(
+                f"   🌟 Best Patch: #{best.patch_id} at {best.coordinates} (score: {best.score:.4f})"
+            )
+
             # Show top patches
-            print(f"   📍 Top Patches:")
+            print("   📍 Top Patches:")
             for j, patch in enumerate(result.all_patches[:max_patches_per_doc]):
-                print(f"      {j+1}. Patch #{patch.patch_id}: {patch.score:.4f} at {patch.coordinates}")
-            
+                print(
+                    f"      {j + 1}. Patch #{patch.patch_id}: {patch.score:.4f} at {patch.coordinates}"
+                )
+
             # Show spatial clusters if available
             if result.spatial_clusters and len(result.spatial_clusters) > 1:
                 print(f"   🗂️ Spatial Clusters: {len(result.spatial_clusters)}")
                 for j, cluster in enumerate(result.spatial_clusters[:2]):  # Show top 2 clusters
                     cluster_score = max(p.score for p in cluster)
-                    print(f"      Cluster {j+1}: {len(cluster)} patches (best: {cluster_score:.4f})")
+                    print(
+                        f"      Cluster {j + 1}: {len(cluster)} patches (best: {cluster_score:.4f})"
+                    )
+
 
 def demo_aggregation():
     """Demonstrate the multi-vector aggregation functionality."""
     print("=== Multi-Vector Aggregation Demo ===")
-    
+
     # Simulate some patch-level search results
     # In real usage, these would come from LeannSearcher.search()
-    
+
     class MockResult:
         def __init__(self, score, metadata):
             self.score = score
             self.metadata = metadata
-    
+
     # Simulate results for 2 images with multiple patches each
     mock_results = [
         # Image 1: cats_and_kitchen.jpg - 4 patches
-        MockResult(0.85, {
-            "image_name": "cats_and_kitchen.jpg",
-            "image_path": "/path/to/cats_and_kitchen.jpg",
-            "patch_id": 3,
-            "coordinates": [100, 50, 224, 174],  # Kitchen area
-            "attention_score": 0.92,
-            "scale": 1.0
-        }),
-        MockResult(0.78, {
-            "image_name": "cats_and_kitchen.jpg", 
-            "image_path": "/path/to/cats_and_kitchen.jpg",
-            "patch_id": 7,
-            "coordinates": [200, 300, 324, 424],  # Cat area
-            "attention_score": 0.88,
-            "scale": 1.0
-        }),
-        MockResult(0.72, {
-            "image_name": "cats_and_kitchen.jpg",
-            "image_path": "/path/to/cats_and_kitchen.jpg", 
-            "patch_id": 12,
-            "coordinates": [150, 100, 274, 224],  # Appliances
-            "attention_score": 0.75,
-            "scale": 1.0
-        }),
-        MockResult(0.65, {
-            "image_name": "cats_and_kitchen.jpg",
-            "image_path": "/path/to/cats_and_kitchen.jpg",
-            "patch_id": 15,
-            "coordinates": [50, 250, 174, 374],  # Furniture
-            "attention_score": 0.70,
-            "scale": 1.0
-        }),
-        
-        # Image 2: city_street.jpg - 3 patches  
-        MockResult(0.68, {
-            "image_name": "city_street.jpg",
-            "image_path": "/path/to/city_street.jpg",
-            "patch_id": 2,
-            "coordinates": [300, 100, 424, 224],  # Buildings
-            "attention_score": 0.80,
-            "scale": 1.0
-        }),
-        MockResult(0.62, {
-            "image_name": "city_street.jpg",
-            "image_path": "/path/to/city_street.jpg",
-            "patch_id": 8,
-            "coordinates": [100, 350, 224, 474],  # Street level
-            "attention_score": 0.75,
-            "scale": 1.0
-        }),
-        MockResult(0.55, {
-            "image_name": "city_street.jpg", 
-            "image_path": "/path/to/city_street.jpg",
-            "patch_id": 11,
-            "coordinates": [400, 200, 524, 324],  # Sky area
-            "attention_score": 0.60,
-            "scale": 1.0
-        }),
+        MockResult(
+            0.85,
+            {
+                "image_name": "cats_and_kitchen.jpg",
+                "image_path": "/path/to/cats_and_kitchen.jpg",
+                "patch_id": 3,
+                "coordinates": [100, 50, 224, 174],  # Kitchen area
+                "attention_score": 0.92,
+                "scale": 1.0,
+            },
+        ),
+        MockResult(
+            0.78,
+            {
+                "image_name": "cats_and_kitchen.jpg",
+                "image_path": "/path/to/cats_and_kitchen.jpg",
+                "patch_id": 7,
+                "coordinates": [200, 300, 324, 424],  # Cat area
+                "attention_score": 0.88,
+                "scale": 1.0,
+            },
+        ),
+        MockResult(
+            0.72,
+            {
+                "image_name": "cats_and_kitchen.jpg",
+                "image_path": "/path/to/cats_and_kitchen.jpg",
+                "patch_id": 12,
+                "coordinates": [150, 100, 274, 224],  # Appliances
+                "attention_score": 0.75,
+                "scale": 1.0,
+            },
+        ),
+        MockResult(
+            0.65,
+            {
+                "image_name": "cats_and_kitchen.jpg",
+                "image_path": "/path/to/cats_and_kitchen.jpg",
+                "patch_id": 15,
+                "coordinates": [50, 250, 174, 374],  # Furniture
+                "attention_score": 0.70,
+                "scale": 1.0,
+            },
+        ),
+        # Image 2: city_street.jpg - 3 patches
+        MockResult(
+            0.68,
+            {
+                "image_name": "city_street.jpg",
+                "image_path": "/path/to/city_street.jpg",
+                "patch_id": 2,
+                "coordinates": [300, 100, 424, 224],  # Buildings
+                "attention_score": 0.80,
+                "scale": 1.0,
+            },
+        ),
+        MockResult(
+            0.62,
+            {
+                "image_name": "city_street.jpg",
+                "image_path": "/path/to/city_street.jpg",
+                "patch_id": 8,
+                "coordinates": [100, 350, 224, 474],  # Street level
+                "attention_score": 0.75,
+                "scale": 1.0,
+            },
+        ),
+        MockResult(
+            0.55,
+            {
+                "image_name": "city_street.jpg",
+                "image_path": "/path/to/city_street.jpg",
+                "patch_id": 11,
+                "coordinates": [400, 200, 524, 324],  # Sky area
+                "attention_score": 0.60,
+                "scale": 1.0,
+            },
+        ),
     ]
-    
+
     # Test different aggregation methods
     methods = ["maxsim", "voting", "weighted", "mean"]
-    
+
     for method in methods:
-        print(f"\n{'='*20} {method.upper()} AGGREGATION {'='*20}")
-        
+        print(f"\n{'=' * 20} {method.upper()} AGGREGATION {'=' * 20}")
+
         aggregator = MultiVectorAggregator(
             aggregation_method=method,
             spatial_clustering=True,
-            cluster_distance_threshold=100.0
+            cluster_distance_threshold=100.0,
         )
-        
+
         aggregated = aggregator.aggregate_results(mock_results, top_k=5)
         aggregator.print_aggregated_results(aggregated)
 
+
 if __name__ == "__main__":
-    demo_aggregation()
+    demo_aggregation()

examples/openai_hnsw_example.py

@@ -6,22 +6,24 @@ Complete example showing how to build and search with OpenAI embeddings using HN
 """
 
 import os
-import dotenv
 from pathlib import Path
+
+import dotenv
 from leann.api import LeannBuilder, LeannSearcher
 
 # Load environment variables
 dotenv.load_dotenv()
 
+
 def main():
     # Check if OpenAI API key is available
     api_key = os.getenv("OPENAI_API_KEY")
     if not api_key:
         print("ERROR: OPENAI_API_KEY environment variable not set")
         return False
-    
+
     print(f"✅ OpenAI API key found: {api_key[:10]}...")
-    
+
     # Sample texts
     sample_texts = [
         "Machine learning is a powerful technology that enables computers to learn from data.",
@@ -33,15 +35,15 @@ def main():
         "Artificial intelligence aims to create machines that can perform human-like tasks.",
         "Python is a popular programming language used extensively in data science and AI.",
         "Neural networks are inspired by the structure and function of the human brain.",
-        "Big data refers to extremely large datasets that require special tools to process."
+        "Big data refers to extremely large datasets that require special tools to process.",
     ]
-    
+
     INDEX_DIR = Path("./simple_openai_test_index")
     INDEX_PATH = str(INDEX_DIR / "simple_test.leann")
-    
-    print(f"\n=== Building Index with OpenAI Embeddings ===")
+
+    print("\n=== Building Index with OpenAI Embeddings ===")
     print(f"Index path: {INDEX_PATH}")
-    
+
     try:
         # Use proper configuration for OpenAI embeddings
         builder = LeannBuilder(
@@ -49,60 +51,63 @@ def main():
             embedding_model="text-embedding-3-small",
             embedding_mode="openai",
             # HNSW settings for OpenAI embeddings
-            M=16,                    # Smaller graph degree
-            efConstruction=64,       # Smaller construction complexity  
-            is_compact=True,         # Enable compact storage for recompute
-            is_recompute=True,       # MUST enable for OpenAI embeddings
+            M=16,  # Smaller graph degree
+            efConstruction=64,  # Smaller construction complexity
+            is_compact=True,  # Enable compact storage for recompute
+            is_recompute=True,  # MUST enable for OpenAI embeddings
             num_threads=1,
         )
-        
+
         print(f"Adding {len(sample_texts)} texts to the index...")
         for i, text in enumerate(sample_texts):
             metadata = {"id": f"doc_{i}", "topic": "AI"}
             builder.add_text(text, metadata)
-        
+
         print("Building index...")
         builder.build_index(INDEX_PATH)
-        print(f"✅ Index built successfully!")
-        
+        print("✅ Index built successfully!")
+
     except Exception as e:
         print(f"❌ Error building index: {e}")
         import traceback
+
         traceback.print_exc()
         return False
-    
-    print(f"\n=== Testing Search ===")
-    
+
+    print("\n=== Testing Search ===")
+
     try:
         searcher = LeannSearcher(INDEX_PATH)
-        
+
         test_queries = [
             "What is machine learning?",
             "How do neural networks work?",
-            "Programming languages for data science"
+            "Programming languages for data science",
         ]
-        
+
         for query in test_queries:
             print(f"\n🔍 Query: '{query}'")
             results = searcher.search(query, top_k=3)
-            
+
             print(f"   Found {len(results)} results:")
             for i, result in enumerate(results):
-                print(f"   {i+1}. Score: {result.score:.4f}")
+                print(f"   {i + 1}. Score: {result.score:.4f}")
                 print(f"      Text: {result.text[:80]}...")
-        
-        print(f"\n✅ Search test completed successfully!")
+
+        print("\n✅ Search test completed successfully!")
         return True
-        
+
     except Exception as e:
         print(f"❌ Error during search: {e}")
         import traceback
+
         traceback.print_exc()
         return False
 
+
 if __name__ == "__main__":
     success = main()
     if success:
-        print(f"\n🎉 Simple OpenAI index test completed successfully!")
+        print("\n🎉 Simple OpenAI index test completed successfully!")
     else:
-        print(f"\n💥 Simple OpenAI index test failed!")
+        print("\n💥 Simple OpenAI index test failed!")

examples/resue_index.py

@@ -1,18 +1,23 @@
 import asyncio
-from leann.api import LeannChat
 from pathlib import Path
 
+from leann.api import LeannChat
+
 INDEX_DIR = Path("./test_pdf_index_huawei")
 INDEX_PATH = str(INDEX_DIR / "pdf_documents.leann")
 
+
 async def main():
-    print(f"\n[PHASE 2] Starting Leann chat session...")
+    print("\n[PHASE 2] Starting Leann chat session...")
     chat = LeannChat(index_path=INDEX_PATH)
     query = "What is the main idea of RL and give me 5 exapmle of classic RL algorithms?"
     query = "Based on the paper, what are the main techniques LEANN explores to reduce the storage overhead and DLPM explore to achieve Fairness and Efiiciency trade-off?"
-    # query = "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"
-    response = chat.ask(query,top_k=20,recompute_beighbor_embeddings=True,complexity=32,beam_width=1)
+    # query = "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"
+    response = chat.ask(
+        query, top_k=20, recompute_beighbor_embeddings=True, complexity=32, beam_width=1
+    )
     print(f"\n[PHASE 2] Response: {response}")
 
+
 if __name__ == "__main__":
-    asyncio.run(main())
+    asyncio.run(main())

examples/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, 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
 
@@ -63,7 +60,7 @@ def download_data_if_needed(data_root: Path, download_embeddings: bool = False):
             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,
@@ -202,9 +193,7 @@ 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."
     )
@@ -219,9 +208,7 @@ def main():
     # 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 +249,7 @@ 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
@@ -293,12 +278,8 @@ def main():
                         break
 
             if not args.index_path:
-                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."
-                )
+                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.")
                 sys.exit(1)
 
     # Detect dataset type from index path to select the correct ground truth
@@ -310,14 +291,10 @@ 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 +304,7 @@ 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))
@@ -339,9 +316,7 @@ 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
-            )
+            new_results = searcher.search(queries[i], top_k=args.top_k, ef=args.ef_search)
             search_times.append(time.time() - start_time)
 
             # Correct Recall Calculation: Based on TEXT content

examples/simple_demo.py

@@ -4,18 +4,25 @@ Run: uv run python examples/simple_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")
 
 
 if __name__ == "__main__":
-    main()
+    main()

examples/wechat_history_reader_leann.py

@@ -1,13 +1,11 @@
-import os
-import asyncio
-import dotenv
 import argparse
+import asyncio
+import os
 from pathlib import Path
-from typing import List, Any, Optional
-from leann.api import LeannBuilder, LeannSearcher, LeannChat
+
+import dotenv
+from leann.api import LeannBuilder, LeannChat
 from llama_index.core.node_parser import SentenceSplitter
-import requests
-import time
 
 dotenv.load_dotenv()
 
@@ -16,7 +14,7 @@ DEFAULT_WECHAT_EXPORT_DIR = "./wechat_export_direct"
 
 
 def create_leann_index_from_multiple_wechat_exports(
-    export_dirs: List[Path],
+    export_dirs: list[Path],
     index_path: str = "wechat_history_index.leann",
     max_count: int = -1,
 ):
@@ -38,21 +36,19 @@ def create_leann_index_from_multiple_wechat_exports(
     INDEX_DIR = Path(index_path).parent
 
     if not INDEX_DIR.exists():
-        print(f"--- Index directory not found, building new index ---")
+        print("--- Index directory not found, building new index ---")
         all_documents = []
         total_processed = 0
 
         # Process each WeChat export directory
         for i, export_dir in enumerate(export_dirs):
-            print(
-                f"\nProcessing WeChat export {i + 1}/{len(export_dirs)}: {export_dir}"
-            )
+            print(f"\nProcessing WeChat export {i + 1}/{len(export_dirs)}: {export_dir}")
 
             try:
                 documents = reader.load_data(
                     wechat_export_dir=str(export_dir),
                     max_count=max_count,
-                    concatenate_messages=False,  # Disable concatenation - one message per document
+                    concatenate_messages=True,  # Disable concatenation - one message per document
                 )
                 if documents:
                     print(f"Loaded {len(documents)} chat documents from {export_dir}")
@@ -74,11 +70,11 @@ def create_leann_index_from_multiple_wechat_exports(
             return None
 
         print(
-            f"\nTotal loaded {len(all_documents)} chat documents from {len(export_dirs)} exports"
+            f"\nTotal loaded {len(all_documents)} chat documents from {len(export_dirs)} exports and starting to split them into chunks"
         )
 
         # Create text splitter with 256 chunk size
-        text_splitter = SentenceSplitter(chunk_size=128, chunk_overlap=64)
+        text_splitter = SentenceSplitter(chunk_size=192, chunk_overlap=64)
 
         # Convert Documents to text strings and chunk them
         all_texts = []
@@ -86,19 +82,25 @@ def create_leann_index_from_multiple_wechat_exports(
             # Split the document into chunks
             nodes = text_splitter.get_nodes_from_documents([doc])
             for node in nodes:
-                all_texts.append(node.get_content())
+                text = (
+                    "[Contact] means the message is from: "
+                    + doc.metadata["contact_name"]
+                    + "\n"
+                    + node.get_content()
+                )
+                all_texts.append(text)
 
         print(
-            f"Created {len(all_texts)} text chunks from {len(all_documents)} documents"
+            f"Finished splitting {len(all_documents)} documents into {len(all_texts)} text chunks"
         )
 
         # Create LEANN index directory
-        print(f"--- Index directory not found, building new index ---")
+        print("--- Index directory not found, building new index ---")
         INDEX_DIR.mkdir(exist_ok=True)
 
-        print(f"--- Building new LEANN index ---")
+        print("--- Building new LEANN index ---")
 
-        print(f"\n[PHASE 1] Building Leann index...")
+        print("\n[PHASE 1] Building Leann index...")
 
         # Use HNSW backend for better macOS compatibility
         builder = LeannBuilder(
@@ -124,7 +126,7 @@ def create_leann_index_from_multiple_wechat_exports(
 
 
 def create_leann_index(
-    export_dir: str = None,
+    export_dir: str | None = None,
     index_path: str = "wechat_history_index.leann",
     max_count: int = 1000,
 ):
@@ -140,12 +142,12 @@ def create_leann_index(
     INDEX_DIR = Path(index_path).parent
 
     if not INDEX_DIR.exists():
-        print(f"--- Index directory not found, building new index ---")
+        print("--- Index directory not found, building new index ---")
         INDEX_DIR.mkdir(exist_ok=True)
 
-        print(f"--- Building new LEANN index ---")
+        print("--- Building new LEANN index ---")
 
-        print(f"\n[PHASE 1] Building Leann index...")
+        print("\n[PHASE 1] Building Leann index...")
 
         # Load documents using WeChatHistoryReader from history_data
         from history_data.wechat_history import WeChatHistoryReader
@@ -178,12 +180,12 @@ def create_leann_index(
         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 ---")
+        print("--- Index directory not found, building new index ---")
         INDEX_DIR.mkdir(exist_ok=True)
 
-        print(f"--- Building new LEANN index ---")
+        print("--- Building new LEANN index ---")
 
-        print(f"\n[PHASE 1] Building Leann index...")
+        print("\n[PHASE 1] Building Leann index...")
 
         # Use HNSW backend for better macOS compatibility
         builder = LeannBuilder(
@@ -216,7 +218,7 @@ async def query_leann_index(index_path: str, query: str):
         index_path: Path to the LEANN index
         query: The query string
     """
-    print(f"\n[PHASE 2] Starting Leann chat session...")
+    print("\n[PHASE 2] Starting Leann chat session...")
     chat = LeannChat(index_path=index_path)
 
     print(f"You: {query}")
@@ -224,7 +226,7 @@ async def query_leann_index(index_path: str, query: str):
         query,
         top_k=20,
         recompute_beighbor_embeddings=True,
-        complexity=128,
+        complexity=16,
         beam_width=1,
         llm_config={
             "type": "openai",
@@ -233,7 +235,7 @@ async def query_leann_index(index_path: str, query: str):
         },
         llm_kwargs={"temperature": 0.0, "max_tokens": 1000},
     )
-    print(f"Leann: {chat_response}")
+    print(f"Leann chat response: \033[36m{chat_response}\033[0m")
 
 
 async def main():
@@ -252,13 +254,13 @@ async def main():
     parser.add_argument(
         "--index-dir",
         type=str,
-        default="./wechat_history_june19_test",
+        default="./wechat_history_magic_test_11Debug_new",
         help="Directory to store the LEANN index (default: ./wechat_history_index_leann_test)",
     )
     parser.add_argument(
         "--max-entries",
         type=int,
-        default=5000,
+        default=50,
         help="Maximum number of chat entries to process (default: 5000)",
     )
     parser.add_argument(
@@ -306,7 +308,7 @@ async def main():
         else:
             # Example queries
             queries = [
-                "我想买魔术师约翰逊的球衣，给我一些对应聊天记录?",
+                "我想买魔术师约翰逊的球衣,给我一些对应聊天记录?",
             ]
 
             for query in queries:

packages/__init__.py

@@ -1 +0,0 @@
-

packages/leann-backend-diskann/__init__.py

@@ -1 +1 @@
-# This file makes the directory a Python package 
+# This file makes the directory a Python package

packages/leann-backend-diskann/leann_backend_diskann/__init__.py

@@ -1 +1 @@
-from . import diskann_backend
+from . import diskann_backend as diskann_backend

packages/leann-backend-diskann/leann_backend_diskann/diskann_backend.py

@@ -1,18 +1,60 @@
-import numpy as np
+import contextlib
+import logging
 import os
 import struct
+import sys
 from pathlib import Path
-from typing import Dict, Any, List, Literal
-import contextlib
-import pickle
+from typing import Any, Literal
 
-from leann.searcher_base import BaseSearcher
-from leann.registry import register_backend
+import numpy as np
+import psutil
 from leann.interface import (
-    LeannBackendFactoryInterface,
     LeannBackendBuilderInterface,
+    LeannBackendFactoryInterface,
     LeannBackendSearcherInterface,
 )
+from leann.registry import register_backend
+from leann.searcher_base import BaseSearcher
+
+logger = logging.getLogger(__name__)
+
+
+@contextlib.contextmanager
+def suppress_cpp_output_if_needed():
+    """Suppress C++ stdout/stderr based on LEANN_LOG_LEVEL"""
+    log_level = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
+
+    # Only suppress if log level is WARNING or higher (ERROR, CRITICAL)
+    should_suppress = log_level in ["WARNING", "ERROR", "CRITICAL"]
+
+    if not should_suppress:
+        # Don't suppress, just yield
+        yield
+        return
+
+    # Save original file descriptors
+    stdout_fd = sys.stdout.fileno()
+    stderr_fd = sys.stderr.fileno()
+
+    # Save original stdout/stderr
+    stdout_dup = os.dup(stdout_fd)
+    stderr_dup = os.dup(stderr_fd)
+
+    try:
+        # Redirect to /dev/null
+        devnull = os.open(os.devnull, os.O_WRONLY)
+        os.dup2(devnull, stdout_fd)
+        os.dup2(devnull, stderr_fd)
+        os.close(devnull)
+
+        yield
+
+    finally:
+        # Restore original file descriptors
+        os.dup2(stdout_dup, stdout_fd)
+        os.dup2(stderr_dup, stderr_fd)
+        os.close(stdout_dup)
+        os.close(stderr_dup)
 
 
 def _get_diskann_metrics():
@@ -43,6 +85,43 @@ def _write_vectors_to_bin(data: np.ndarray, file_path: Path):
         f.write(data.tobytes())
 
 
+def _calculate_smart_memory_config(data: np.ndarray) -> tuple[float, float]:
+    """
+    Calculate smart memory configuration for DiskANN based on data size and system specs.
+
+    Args:
+        data: The embedding data array
+
+    Returns:
+        tuple: (search_memory_maximum, build_memory_maximum) in GB
+    """
+    num_vectors, dim = data.shape
+
+    # Calculate embedding storage size
+    embedding_size_bytes = num_vectors * dim * 4  # float32 = 4 bytes
+    embedding_size_gb = embedding_size_bytes / (1024**3)
+
+    # search_memory_maximum: 1/10 of embedding size for optimal PQ compression
+    # This controls Product Quantization size - smaller means more compression
+    search_memory_gb = max(0.1, embedding_size_gb / 10)  # At least 100MB
+
+    # build_memory_maximum: Based on available system RAM for sharding control
+    # This controls how much memory DiskANN uses during index construction
+    available_memory_gb = psutil.virtual_memory().available / (1024**3)
+    total_memory_gb = psutil.virtual_memory().total / (1024**3)
+
+    # Use 50% of available memory, but at least 2GB and at most 75% of total
+    build_memory_gb = max(2.0, min(available_memory_gb * 0.5, total_memory_gb * 0.75))
+
+    logger.info(
+        f"Smart memory config - Data: {embedding_size_gb:.2f}GB, "
+        f"Search mem: {search_memory_gb:.2f}GB (PQ control), "
+        f"Build mem: {build_memory_gb:.2f}GB (sharding control)"
+    )
+
+    return search_memory_gb, build_memory_gb
+
+
 @register_backend("diskann")
 class DiskannBackend(LeannBackendFactoryInterface):
     @staticmethod
@@ -58,25 +137,37 @@ class DiskannBuilder(LeannBackendBuilderInterface):
     def __init__(self, **kwargs):
         self.build_params = kwargs
 
-    def build(self, data: np.ndarray, ids: List[str], index_path: str, **kwargs):
+    def build(self, data: np.ndarray, ids: list[str], index_path: str, **kwargs):
         path = Path(index_path)
         index_dir = path.parent
         index_prefix = path.stem
         index_dir.mkdir(parents=True, exist_ok=True)
 
         if data.dtype != np.float32:
+            logger.warning(f"Converting data to float32, shape: {data.shape}")
             data = data.astype(np.float32)
 
         data_filename = f"{index_prefix}_data.bin"
         _write_vectors_to_bin(data, index_dir / data_filename)
 
-
         build_kwargs = {**self.build_params, **kwargs}
         metric_enum = _get_diskann_metrics().get(
             build_kwargs.get("distance_metric", "mips").lower()
         )
         if metric_enum is None:
-            raise ValueError("Unsupported distance_metric.")
+            raise ValueError(
+                f"Unsupported distance_metric '{build_kwargs.get('distance_metric', 'unknown')}'."
+            )
+
+        # Calculate smart memory configuration if not explicitly provided
+        if (
+            "search_memory_maximum" not in build_kwargs
+            or "build_memory_maximum" not in build_kwargs
+        ):
+            smart_search_mem, smart_build_mem = _calculate_smart_memory_config(data)
+        else:
+            smart_search_mem = build_kwargs.get("search_memory_maximum", 4.0)
+            smart_build_mem = build_kwargs.get("build_memory_maximum", 8.0)
 
         try:
             from . import _diskannpy as diskannpy  # type: ignore
@@ -88,8 +179,8 @@ class DiskannBuilder(LeannBackendBuilderInterface):
                     index_prefix,
                     build_kwargs.get("complexity", 64),
                     build_kwargs.get("graph_degree", 32),
-                    build_kwargs.get("search_memory_maximum", 4.0),
-                    build_kwargs.get("build_memory_maximum", 8.0),
+                    build_kwargs.get("search_memory_maximum", smart_search_mem),
+                    build_kwargs.get("build_memory_maximum", smart_build_mem),
                     build_kwargs.get("num_threads", 8),
                     build_kwargs.get("pq_disk_bytes", 0),
                     "",
@@ -98,36 +189,66 @@ class DiskannBuilder(LeannBackendBuilderInterface):
             temp_data_file = index_dir / data_filename
             if temp_data_file.exists():
                 os.remove(temp_data_file)
+                logger.debug(f"Cleaned up temporary data file: {temp_data_file}")
 
 
 class DiskannSearcher(BaseSearcher):
     def __init__(self, index_path: str, **kwargs):
         super().__init__(
             index_path,
-            backend_module_name="leann_backend_diskann.embedding_server",
+            backend_module_name="leann_backend_diskann.diskann_embedding_server",
             **kwargs,
         )
-        from . import _diskannpy as diskannpy  # type: ignore
 
-        distance_metric = kwargs.get("distance_metric", "mips").lower()
-        metric_enum = _get_diskann_metrics().get(distance_metric)
-        if metric_enum is None:
-            raise ValueError(f"Unsupported distance_metric '{distance_metric}'.")
+        # Initialize DiskANN index with suppressed C++ output based on log level
+        with suppress_cpp_output_if_needed():
+            from . import _diskannpy as diskannpy  # type: ignore
 
-        self.num_threads = kwargs.get("num_threads", 8)
-        self.zmq_port = kwargs.get("zmq_port", 6666)
+            distance_metric = kwargs.get("distance_metric", "mips").lower()
+            metric_enum = _get_diskann_metrics().get(distance_metric)
+            if metric_enum is None:
+                raise ValueError(f"Unsupported distance_metric '{distance_metric}'.")
 
-        full_index_prefix = str(self.index_dir / self.index_path.stem)
-        self._index = diskannpy.StaticDiskFloatIndex(
-            metric_enum,
-            full_index_prefix,
-            self.num_threads,
-            kwargs.get("num_nodes_to_cache", 0),
-            1,
-            self.zmq_port,
-            "",
-            "",
-        )
+            self.num_threads = kwargs.get("num_threads", 8)
+
+            # For DiskANN, we need to reinitialize the index when zmq_port changes
+            # Store the initialization parameters for later use
+            full_index_prefix = str(self.index_dir / self.index_path.stem)
+            self._init_params = {
+                "metric_enum": metric_enum,
+                "full_index_prefix": full_index_prefix,
+                "num_threads": self.num_threads,
+                "num_nodes_to_cache": kwargs.get("num_nodes_to_cache", 0),
+                "cache_mechanism": 1,
+                "pq_prefix": "",
+                "partition_prefix": "",
+            }
+            self._diskannpy = diskannpy
+            self._current_zmq_port = None
+            self._index = None
+            logger.debug("DiskANN searcher initialized (index will be loaded on first search)")
+
+    def _ensure_index_loaded(self, zmq_port: int):
+        """Ensure the index is loaded with the correct zmq_port."""
+        if self._index is None or self._current_zmq_port != zmq_port:
+            # Need to (re)load the index with the correct zmq_port
+            with suppress_cpp_output_if_needed():
+                if self._index is not None:
+                    logger.debug(f"Reloading DiskANN index with new zmq_port: {zmq_port}")
+                else:
+                    logger.debug(f"Loading DiskANN index with zmq_port: {zmq_port}")
+
+                self._index = self._diskannpy.StaticDiskFloatIndex(
+                    self._init_params["metric_enum"],
+                    self._init_params["full_index_prefix"],
+                    self._init_params["num_threads"],
+                    self._init_params["num_nodes_to_cache"],
+                    self._init_params["cache_mechanism"],
+                    zmq_port,
+                    self._init_params["pq_prefix"],
+                    self._init_params["partition_prefix"],
+                )
+                self._current_zmq_port = zmq_port
 
     def search(
         self,
@@ -138,11 +259,11 @@ class DiskannSearcher(BaseSearcher):
         prune_ratio: float = 0.0,
         recompute_embeddings: bool = False,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: int = 5557,
+        zmq_port: int | None = None,
         batch_recompute: bool = False,
         dedup_node_dis: bool = False,
         **kwargs,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         """
         Search for nearest neighbors using DiskANN index.
 
@@ -157,7 +278,7 @@ class DiskannSearcher(BaseSearcher):
                 - "global": Use global pruning strategy (default)
                 - "local": Use local pruning strategy
                 - "proportional": Not supported in DiskANN, falls back to global
-            zmq_port: ZMQ port for embedding server
+            zmq_port: ZMQ port for embedding server communication. Must be provided if recompute_embeddings is True.
             batch_recompute: Whether to batch neighbor recomputation (DiskANN-specific)
             dedup_node_dis: Whether to cache and reuse distance computations (DiskANN-specific)
             **kwargs: Additional DiskANN-specific parameters (for legacy compatibility)
@@ -165,22 +286,22 @@ class DiskannSearcher(BaseSearcher):
         Returns:
             Dict with 'labels' (list of lists) and 'distances' (ndarray)
         """
+        # Handle zmq_port compatibility: Ensure index is loaded with correct port
+        if recompute_embeddings:
+            if zmq_port is None:
+                raise ValueError("zmq_port must be provided if recompute_embeddings is True")
+            self._ensure_index_loaded(zmq_port)
+        else:
+            # If not recomputing, we still need an index, use a default port
+            if self._index is None:
+                self._ensure_index_loaded(6666)  # Default port when not recomputing
+
         # DiskANN doesn't support "proportional" strategy
         if pruning_strategy == "proportional":
             raise NotImplementedError(
                 "DiskANN backend does not support 'proportional' pruning strategy. Use 'global' or 'local' instead."
             )
 
-        # Use recompute_embeddings parameter
-        use_recompute = recompute_embeddings
-        if use_recompute:
-            meta_file_path = self.index_dir / f"{self.index_path.name}.meta.json"
-            if not meta_file_path.exists():
-                raise RuntimeError(
-                    f"FATAL: Recompute enabled but metadata file not found: {meta_file_path}"
-                )
-            self._ensure_server_running(str(meta_file_path), port=zmq_port, **kwargs)
-
         if query.dtype != np.float32:
             query = query.astype(np.float32)
 
@@ -190,25 +311,24 @@ class DiskannSearcher(BaseSearcher):
         else:  # "global"
             use_global_pruning = True
 
-        labels, distances = self._index.batch_search(
-            query,
-            query.shape[0],
-            top_k,
-            complexity,
-            beam_width,
-            self.num_threads,
-            kwargs.get("USE_DEFERRED_FETCH", False),
-            kwargs.get("skip_search_reorder", False),
-            use_recompute,
-            dedup_node_dis,
-            prune_ratio,
-            batch_recompute,
-            use_global_pruning,
-        )
+        # Perform search with suppressed C++ output based on log level
+        with suppress_cpp_output_if_needed():
+            labels, distances = self._index.batch_search(
+                query,
+                query.shape[0],
+                top_k,
+                complexity,
+                beam_width,
+                self.num_threads,
+                kwargs.get("USE_DEFERRED_FETCH", False),
+                kwargs.get("skip_search_reorder", False),
+                recompute_embeddings,
+                dedup_node_dis,
+                prune_ratio,
+                batch_recompute,
+                use_global_pruning,
+            )
 
-        string_labels = [
-            [str(int_label) for int_label in batch_labels]
-            for batch_labels in labels
-        ]
+        string_labels = [[str(int_label) for int_label in batch_labels] for batch_labels in labels]
 
         return {"labels": string_labels, "distances": distances}

packages/leann-backend-diskann/leann_backend_diskann/diskann_embedding_server.py

@@ -0,0 +1,284 @@
+"""
+DiskANN-specific embedding server
+"""
+
+import argparse
+import json
+import logging
+import os
+import sys
+import threading
+import time
+from pathlib import Path
+
+import numpy as np
+import zmq
+
+# Set up logging based on environment variable
+LOG_LEVEL = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
+logger = logging.getLogger(__name__)
+
+# Force set logger level (don't rely on basicConfig in subprocess)
+log_level = getattr(logging, LOG_LEVEL, logging.WARNING)
+logger.setLevel(log_level)
+
+# Ensure we have a handler if none exists
+if not logger.handlers:
+    handler = logging.StreamHandler()
+    formatter = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
+    handler.setFormatter(formatter)
+    logger.addHandler(handler)
+    logger.propagate = False
+
+
+def create_diskann_embedding_server(
+    passages_file: str | None = None,
+    zmq_port: int = 5555,
+    model_name: str = "sentence-transformers/all-mpnet-base-v2",
+    embedding_mode: str = "sentence-transformers",
+    distance_metric: str = "l2",
+):
+    """
+    Create and start a ZMQ-based embedding server for DiskANN backend.
+    Uses ROUTER socket and protobuf communication as required by DiskANN C++ implementation.
+    """
+    logger.info(f"Starting DiskANN server on port {zmq_port} with model {model_name}")
+    logger.info(f"Using embedding mode: {embedding_mode}")
+
+    # Add leann-core to path for unified embedding computation
+    current_dir = Path(__file__).parent
+    leann_core_path = current_dir.parent.parent / "leann-core" / "src"
+    sys.path.insert(0, str(leann_core_path))
+
+    try:
+        from leann.api import PassageManager
+        from leann.embedding_compute import compute_embeddings
+
+        logger.info("Successfully imported unified embedding computation module")
+    except ImportError as e:
+        logger.error(f"Failed to import embedding computation module: {e}")
+        return
+    finally:
+        sys.path.pop(0)
+
+    # Check port availability
+    import socket
+
+    def check_port(port):
+        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+            return s.connect_ex(("localhost", port)) == 0
+
+    if check_port(zmq_port):
+        logger.error(f"Port {zmq_port} is already in use")
+        return
+
+    # Only support metadata file, fail fast for everything else
+    if not passages_file or not passages_file.endswith(".meta.json"):
+        raise ValueError("Only metadata files (.meta.json) are supported")
+
+    # Load metadata to get passage sources
+    with open(passages_file) as f:
+        meta = json.load(f)
+
+    passages = PassageManager(meta["passage_sources"])
+    logger.info(
+        f"Loaded PassageManager with {len(passages.global_offset_map)} passages from metadata"
+    )
+
+    # Import protobuf after ensuring the path is correct
+    try:
+        from . import embedding_pb2
+    except ImportError as e:
+        logger.error(f"Failed to import protobuf module: {e}")
+        return
+
+    def zmq_server_thread():
+        """ZMQ server thread using REP socket for universal compatibility"""
+        context = zmq.Context()
+        socket = context.socket(
+            zmq.REP
+        )  # REP socket for both BaseSearcher and DiskANN C++ REQ clients
+        socket.bind(f"tcp://*:{zmq_port}")
+        logger.info(f"DiskANN ZMQ REP server listening on port {zmq_port}")
+
+        socket.setsockopt(zmq.RCVTIMEO, 300000)
+        socket.setsockopt(zmq.SNDTIMEO, 300000)
+
+        while True:
+            try:
+                # REP socket receives single-part messages
+                message = socket.recv()
+
+                # Check for empty messages - REP socket requires response to every request
+                if len(message) == 0:
+                    logger.debug("Received empty message, sending empty response")
+                    socket.send(b"")  # REP socket must respond to every request
+                    continue
+
+                logger.debug(f"Received ZMQ request of size {len(message)} bytes")
+                logger.debug(f"Message preview: {message[:50]}")  # Show first 50 bytes
+
+                e2e_start = time.time()
+
+                # Try protobuf first (for DiskANN C++ node_ids requests - primary use case)
+                texts = []
+                node_ids = []
+                is_text_request = False
+
+                try:
+                    req_proto = embedding_pb2.NodeEmbeddingRequest()
+                    req_proto.ParseFromString(message)
+                    node_ids = list(req_proto.node_ids)
+
+                    if not node_ids:
+                        raise RuntimeError(
+                            f"PROTOBUF: Received empty node_ids! Message size: {len(message)}"
+                        )
+
+                    logger.info(
+                        f"✅ PROTOBUF: Node ID request for {len(node_ids)} node embeddings: {node_ids[:10]}"
+                    )
+                except Exception as protobuf_error:
+                    logger.debug(f"Protobuf parsing failed: {protobuf_error}")
+                    # Fallback to msgpack (for BaseSearcher direct text requests)
+                    try:
+                        import msgpack
+
+                        request = msgpack.unpackb(message)
+                        # For BaseSearcher compatibility, request is a list of texts directly
+                        if isinstance(request, list) and all(
+                            isinstance(item, str) for item in request
+                        ):
+                            texts = request
+                            is_text_request = True
+                            logger.info(f"✅ MSGPACK: Direct text request for {len(texts)} texts")
+                        else:
+                            raise ValueError("Not a valid msgpack text request")
+                    except Exception as msgpack_error:
+                        raise RuntimeError(
+                            f"Both protobuf and msgpack parsing failed! Protobuf: {protobuf_error}, Msgpack: {msgpack_error}"
+                        )
+
+                # Look up texts by node IDs (only if not direct text request)
+                if not is_text_request:
+                    for nid in node_ids:
+                        try:
+                            passage_data = passages.get_passage(str(nid))
+                            txt = passage_data["text"]
+                            if not txt:
+                                raise RuntimeError(f"FATAL: Empty text for passage ID {nid}")
+                            texts.append(txt)
+                        except KeyError as e:
+                            logger.error(f"Passage ID {nid} not found: {e}")
+                            raise e
+                        except Exception as e:
+                            logger.error(f"Exception looking up passage ID {nid}: {e}")
+                            raise
+
+                    # Debug logging
+                    logger.debug(f"Processing {len(texts)} texts")
+                    logger.debug(f"Text lengths: {[len(t) for t in texts[:5]]}")  # Show first 5
+
+                # Process embeddings using unified computation
+                embeddings = compute_embeddings(texts, model_name, mode=embedding_mode)
+                logger.info(
+                    f"Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
+                )
+
+                # Prepare response based on request type
+                if is_text_request:
+                    # For BaseSearcher compatibility: return msgpack format
+                    import msgpack
+
+                    response_data = msgpack.packb(embeddings.tolist())
+                else:
+                    # For DiskANN C++ compatibility: return protobuf format
+                    resp_proto = embedding_pb2.NodeEmbeddingResponse()
+                    hidden_contiguous = np.ascontiguousarray(embeddings, dtype=np.float32)
+
+                    # Serialize embeddings data
+                    resp_proto.embeddings_data = hidden_contiguous.tobytes()
+                    resp_proto.dimensions.append(hidden_contiguous.shape[0])
+                    resp_proto.dimensions.append(hidden_contiguous.shape[1])
+
+                    response_data = resp_proto.SerializeToString()
+
+                # Send response back to the client
+                socket.send(response_data)
+
+                e2e_end = time.time()
+                logger.info(f"⏱️  ZMQ E2E time: {e2e_end - e2e_start:.6f}s")
+
+            except zmq.Again:
+                logger.debug("ZMQ socket timeout, continuing to listen")
+                continue
+            except Exception as e:
+                logger.error(f"Error in ZMQ server loop: {e}")
+                import traceback
+
+                traceback.print_exc()
+                raise
+
+    zmq_thread = threading.Thread(target=zmq_server_thread, daemon=True)
+    zmq_thread.start()
+    logger.info(f"Started DiskANN ZMQ server thread on port {zmq_port}")
+
+    # Keep the main thread alive
+    try:
+        while True:
+            time.sleep(1)
+    except KeyboardInterrupt:
+        logger.info("DiskANN Server shutting down...")
+        return
+
+
+if __name__ == "__main__":
+    import signal
+    import sys
+
+    def signal_handler(sig, frame):
+        logger.info(f"Received signal {sig}, shutting down gracefully...")
+        sys.exit(0)
+
+    # Register signal handlers for graceful shutdown
+    signal.signal(signal.SIGTERM, signal_handler)
+    signal.signal(signal.SIGINT, signal_handler)
+
+    parser = argparse.ArgumentParser(description="DiskANN Embedding service")
+    parser.add_argument("--zmq-port", type=int, default=5555, help="ZMQ port to run on")
+    parser.add_argument(
+        "--passages-file",
+        type=str,
+        help="Metadata JSON file containing passage sources",
+    )
+    parser.add_argument(
+        "--model-name",
+        type=str,
+        default="sentence-transformers/all-mpnet-base-v2",
+        help="Embedding model name",
+    )
+    parser.add_argument(
+        "--embedding-mode",
+        type=str,
+        default="sentence-transformers",
+        choices=["sentence-transformers", "openai", "mlx"],
+        help="Embedding backend mode",
+    )
+    parser.add_argument(
+        "--distance-metric",
+        type=str,
+        default="l2",
+        choices=["l2", "mips", "cosine"],
+        help="Distance metric for similarity computation",
+    )
+
+    args = parser.parse_args()
+
+    # Create and start the DiskANN embedding server
+    create_diskann_embedding_server(
+        passages_file=args.passages_file,
+        zmq_port=args.zmq_port,
+        model_name=args.model_name,
+        embedding_mode=args.embedding_mode,
+        distance_metric=args.distance_metric,
+    )

packages/leann-backend-diskann/leann_backend_diskann/embedding_pb2.py

@@ -1,27 +1,28 @@
-# -*- coding: utf-8 -*-
 # Generated by the protocol buffer compiler.  DO NOT EDIT!
 # source: embedding.proto
+# ruff: noqa
 """Generated protocol buffer code."""
-from google.protobuf.internal import builder as _builder
+
 from google.protobuf import descriptor as _descriptor
 from google.protobuf import descriptor_pool as _descriptor_pool
 from google.protobuf import symbol_database as _symbol_database
+from google.protobuf.internal import builder as _builder
+
 # @@protoc_insertion_point(imports)
 
 _sym_db = _symbol_database.Default()
 
 
-
-
-DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\x0f\x65mbedding.proto\x12\x0eprotoembedding\"(\n\x14NodeEmbeddingRequest\x12\x10\n\x08node_ids\x18\x01 \x03(\r\"Y\n\x15NodeEmbeddingResponse\x12\x17\n\x0f\x65mbeddings_data\x18\x01 \x01(\x0c\x12\x12\n\ndimensions\x18\x02 \x03(\x05\x12\x13\n\x0bmissing_ids\x18\x03 \x03(\rb\x06proto3')
+DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(
+    b'\n\x0f\x65mbedding.proto\x12\x0eprotoembedding"(\n\x14NodeEmbeddingRequest\x12\x10\n\x08node_ids\x18\x01 \x03(\r"Y\n\x15NodeEmbeddingResponse\x12\x17\n\x0f\x65mbeddings_data\x18\x01 \x01(\x0c\x12\x12\n\ndimensions\x18\x02 \x03(\x05\x12\x13\n\x0bmissing_ids\x18\x03 \x03(\rb\x06proto3'
+)
 
 _builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, globals())
-_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, 'embedding_pb2', globals())
-if _descriptor._USE_C_DESCRIPTORS == False:
-
-  DESCRIPTOR._options = None
-  _NODEEMBEDDINGREQUEST._serialized_start=35
-  _NODEEMBEDDINGREQUEST._serialized_end=75
-  _NODEEMBEDDINGRESPONSE._serialized_start=77
-  _NODEEMBEDDINGRESPONSE._serialized_end=166
+_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, "embedding_pb2", globals())
+if not _descriptor._USE_C_DESCRIPTORS:
+    DESCRIPTOR._options = None
+    _NODEEMBEDDINGREQUEST._serialized_start = 35
+    _NODEEMBEDDINGREQUEST._serialized_end = 75
+    _NODEEMBEDDINGRESPONSE._serialized_start = 77
+    _NODEEMBEDDINGRESPONSE._serialized_end = 166
 # @@protoc_insertion_point(module_scope)

packages/leann-backend-diskann/leann_backend_diskann/embedding_server.py

@@ -1,705 +0,0 @@
-#!/usr/bin/env python3
-"""
-Embedding server for leann-backend-diskann - Fixed ZMQ REQ-REP pattern
-"""
-
-import pickle
-import argparse
-import time
-import json
-from typing import Dict, Any, Optional, Union
-
-from transformers import AutoTokenizer, AutoModel
-import os
-from contextlib import contextmanager
-import zmq
-import numpy as np
-import msgpack
-from pathlib import Path
-import logging
-
-RED = "\033[91m"
-
-# Set up logging based on environment variable
-LOG_LEVEL = os.getenv('LEANN_LOG_LEVEL', 'INFO').upper()
-logging.basicConfig(
-    level=getattr(logging, LOG_LEVEL, logging.INFO),
-    format='%(asctime)s - %(levelname)s - %(message)s'
-)
-logger = logging.getLogger(__name__)
-RESET = "\033[0m"
-
-# --- New Passage Loader from HNSW backend ---
-class SimplePassageLoader:
-    """
-    Simple passage loader that replaces config.py dependencies
-    """
-    def __init__(self, passages_data: Optional[Dict[str, Any]] = None):
-        self.passages_data = passages_data or {}
-        self._meta_path = ''
-    
-    def __getitem__(self, passage_id: Union[str, int]) -> Dict[str, str]:
-        """Get passage by ID"""
-        str_id = str(passage_id)
-        if str_id in self.passages_data:
-            return {"text": self.passages_data[str_id]}
-        else:
-            # Return empty text for missing passages
-            return {"text": ""}
-    
-    def __len__(self) -> int:
-        return len(self.passages_data)
-    
-    def keys(self):
-        return self.passages_data.keys()
-
-def load_passages_from_metadata(meta_file: str) -> SimplePassageLoader:
-    """
-    Load passages using metadata file with PassageManager for lazy loading
-    """
-    # Load metadata to get passage sources
-    with open(meta_file, 'r') as f:
-        meta = json.load(f)
-    
-    # Import PassageManager dynamically to avoid circular imports
-    import sys
-    from pathlib import Path
-    
-    # Find the leann package directory relative to this file
-    current_dir = Path(__file__).parent
-    leann_core_path = current_dir.parent.parent / "leann-core" / "src"
-    sys.path.insert(0, str(leann_core_path))
-    
-    try:
-        from leann.api import PassageManager
-        passage_manager = PassageManager(meta['passage_sources'])
-    finally:
-        sys.path.pop(0)
-    
-    print(f"Initialized lazy passage loading for {len(passage_manager.global_offset_map)} passages")
-    
-    class LazyPassageLoader(SimplePassageLoader):
-        def __init__(self, passage_manager):
-            self.passage_manager = passage_manager
-            # Initialize parent with empty data
-            super().__init__({})
-        
-        def __getitem__(self, passage_id: Union[str, int]) -> Dict[str, str]:
-            """Get passage by ID with lazy loading"""
-            try:
-                int_id = int(passage_id)
-                string_id = str(int_id)
-                passage_data = self.passage_manager.get_passage(string_id)
-                if passage_data and passage_data.get("text"):
-                    return {"text": passage_data["text"]}
-                else:
-                    raise RuntimeError(f"FATAL: Empty text for ID {int_id} -> {string_id}")
-            except Exception as e:
-                raise RuntimeError(f"FATAL: Exception getting passage {passage_id}: {e}")
-        
-        def __len__(self) -> int:
-            return len(self.passage_manager.global_offset_map)
-        
-        def keys(self):
-            return self.passage_manager.global_offset_map.keys()
-    
-    loader = LazyPassageLoader(passage_manager)
-    loader._meta_path = meta_file
-    return loader
-
-def load_passages_from_file(passages_file: str) -> SimplePassageLoader:
-    """
-    Load passages from a JSONL file with label map support
-    Expected format: {"id": "passage_id", "text": "passage_text", "metadata": {...}} (one per line)
-    """
-    
-    if not os.path.exists(passages_file):
-        raise FileNotFoundError(f"Passages file {passages_file} not found.")
-    
-    if not passages_file.endswith('.jsonl'):
-        raise ValueError(f"Expected .jsonl file format, got: {passages_file}")
-    
-    # Load passages directly by their sequential IDs
-    passages_data = {}
-    with open(passages_file, 'r', encoding='utf-8') as f:
-        for line in f:
-            if line.strip():
-                passage = json.loads(line)
-                passages_data[passage['id']] = passage['text']
-    
-    print(f"Loaded {len(passages_data)} passages from JSONL file {passages_file}")
-    return SimplePassageLoader(passages_data)
-
-def create_embedding_server_thread(
-    zmq_port=5555,
-    model_name="sentence-transformers/all-mpnet-base-v2",
-    max_batch_size=128,
-    passages_file: Optional[str] = None,
-    embedding_mode: str = "sentence-transformers",
-    enable_warmup: bool = False,
-):
-    """
-    Create and run embedding server in the current thread
-    This function is designed to be called in a separate thread
-    """
-    logger.info(f"Initializing embedding server thread on port {zmq_port}")
-    
-    try:
-        # Check if port is already occupied
-        import socket
-        def check_port(port):
-            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
-                return s.connect_ex(('localhost', port)) == 0
-
-        if check_port(zmq_port):
-            print(f"{RED}Port {zmq_port} is already in use{RESET}")
-            return
-
-        # Auto-detect mode based on model name if not explicitly set
-        if embedding_mode == "sentence-transformers" and model_name.startswith("text-embedding-"):
-            embedding_mode = "openai"
-        
-        if embedding_mode == "mlx":
-            from leann.api import compute_embeddings_mlx
-            import torch
-            logger.info("Using MLX for embeddings")
-            # Set device to CPU for compatibility with DeviceTimer class
-            device = torch.device("cpu")
-            cuda_available = False
-            mps_available = False
-        elif embedding_mode == "openai":
-            from leann.api import compute_embeddings_openai
-            import torch
-            logger.info("Using OpenAI API for embeddings")
-            # Set device to CPU for compatibility with DeviceTimer class
-            device = torch.device("cpu")
-            cuda_available = False
-            mps_available = False
-        elif embedding_mode == "sentence-transformers":
-            # Initialize model
-            tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
-            import torch
-
-            # Select device
-            mps_available = hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()
-            cuda_available = torch.cuda.is_available()
-            
-            if cuda_available:
-                device = torch.device("cuda")
-                logger.info("Using CUDA device")
-            elif mps_available:
-                device = torch.device("mps")
-                logger.info("Using MPS device (Apple Silicon)")
-            else:
-                device = torch.device("cpu")
-                logger.info("Using CPU device")
-            
-            # Load model
-            logger.info(f"Loading model {model_name}")
-            model = AutoModel.from_pretrained(model_name).to(device).eval()
-
-            # Optimize model
-            if cuda_available or mps_available:
-                try:
-                    model = model.half()
-                    model = torch.compile(model)
-                    logger.info(f"Using FP16 precision with model: {model_name}")
-                except Exception as e:
-                    print(f"WARNING: Model optimization failed: {e}")
-        else:
-            raise ValueError(f"Unsupported embedding mode: {embedding_mode}. Supported modes: sentence-transformers, mlx, openai")
-
-        # Load passages from file if provided
-        if passages_file and os.path.exists(passages_file):
-            # Check if it's a metadata file or a single passages file
-            if passages_file.endswith('.meta.json'):
-                passages = load_passages_from_metadata(passages_file)
-            else:
-                # Try to find metadata file in same directory
-                passages_dir = Path(passages_file).parent
-                meta_files = list(passages_dir.glob("*.meta.json"))
-                if meta_files:
-                    print(f"Found metadata file: {meta_files[0]}, using lazy loading")
-                    passages = load_passages_from_metadata(str(meta_files[0]))
-                else:
-                    # Fallback to original single file loading (will cause warnings)
-                    print("WARNING: No metadata file found, using single file loading (may cause missing passage warnings)")
-                    passages = load_passages_from_file(passages_file)
-        else:
-            print("WARNING: No passages file provided or file not found. Using an empty passage loader.")
-            passages = SimplePassageLoader()
-
-        logger.info(f"Loaded {len(passages)} passages.")
-
-        def client_warmup(zmq_port):
-            """Perform client-side warmup for DiskANN server"""
-            time.sleep(2)
-            print(f"Performing client-side warmup with model {model_name}...")
-            
-            # Get actual passage IDs from the loaded passages
-            sample_ids = []
-            if hasattr(passages, 'keys') and len(passages) > 0:
-                available_ids = list(passages.keys())
-                # Take up to 5 actual IDs, but at least 1
-                sample_ids = available_ids[:min(5, len(available_ids))]
-                print(f"Using actual passage IDs for warmup: {sample_ids}")
-            else:
-                print("No passages available for warmup, skipping warmup...")
-                return
-
-            try:
-                context = zmq.Context()
-                socket = context.socket(zmq.REQ)
-                socket.connect(f"tcp://localhost:{zmq_port}")
-                socket.setsockopt(zmq.RCVTIMEO, 30000)
-                socket.setsockopt(zmq.SNDTIMEO, 30000)
-
-                try:
-                    ids_to_send = [int(x) for x in sample_ids]
-                except ValueError:
-                    print("Warning: Could not convert sample IDs to integers, skipping warmup")
-                    return
-
-                if not ids_to_send:
-                    print("Skipping warmup send.")
-                    return
-
-                # Use protobuf format for warmup
-                from . import embedding_pb2
-                req_proto = embedding_pb2.NodeEmbeddingRequest()
-                req_proto.node_ids.extend(ids_to_send)
-                request_bytes = req_proto.SerializeToString()
-
-                for i in range(3):
-                    print(f"Sending warmup request {i + 1}/3 via ZMQ (Protobuf)...")
-                    socket.send(request_bytes)
-                    response_bytes = socket.recv()
-                    
-                    resp_proto = embedding_pb2.NodeEmbeddingResponse()
-                    resp_proto.ParseFromString(response_bytes)
-                    embeddings_count = resp_proto.dimensions[0] if resp_proto.dimensions else 0
-                    print(f"Warmup request {i + 1}/3 successful, received {embeddings_count} embeddings")
-                    time.sleep(0.1)
-
-                print("Client-side Protobuf ZMQ warmup complete")
-                socket.close()
-                context.term()
-            except Exception as e:
-                print(f"Error during Protobuf ZMQ warmup: {e}")
-
-        class DeviceTimer:
-            """Device timer"""
-            def __init__(self, name="", device=device):
-                self.name = name
-                self.device = device
-                self.start_time = 0
-                self.end_time = 0
-                
-                if embedding_mode == "sentence-transformers" and torch.cuda.is_available():
-                    self.start_event = torch.cuda.Event(enable_timing=True)
-                    self.end_event = torch.cuda.Event(enable_timing=True)
-                else:
-                    self.start_event = None
-                    self.end_event = None
-
-            @contextmanager
-            def timing(self):
-                self.start()
-                yield
-                self.end()
-
-            def start(self):
-                if embedding_mode == "sentence-transformers" and torch.cuda.is_available():
-                    torch.cuda.synchronize()
-                    self.start_event.record()
-                else:
-                    if embedding_mode == "sentence-transformers" and self.device.type == "mps":
-                        torch.mps.synchronize()
-                    self.start_time = time.time()
-
-            def end(self):
-                if embedding_mode == "sentence-transformers" and torch.cuda.is_available():
-                    self.end_event.record()
-                    torch.cuda.synchronize()
-                else:
-                    if embedding_mode == "sentence-transformers" and self.device.type == "mps":
-                        torch.mps.synchronize()
-                    self.end_time = time.time()
-
-            def elapsed_time(self):
-                if embedding_mode == "sentence-transformers" and torch.cuda.is_available():
-                    return self.start_event.elapsed_time(self.end_event) / 1000.0
-                else:
-                    return self.end_time - self.start_time
-
-            def print_elapsed(self):
-                elapsed = self.elapsed_time()
-                print(f"[{self.name}] Elapsed time: {elapsed:.3f}s")
-
-        def process_batch_pytorch(texts_batch, ids_batch, missing_ids):
-            """Process text batch"""
-            if not texts_batch:
-                return np.array([])
-
-            # Filter out empty texts and their corresponding IDs
-            valid_texts = []
-            valid_ids = []
-            for i, text in enumerate(texts_batch):
-                if text.strip():  # Only include non-empty texts
-                    valid_texts.append(text)
-                    valid_ids.append(ids_batch[i])
-
-            if not valid_texts:
-                print("WARNING: No valid texts in batch")
-                return np.array([])
-
-            # Tokenize
-            token_timer = DeviceTimer("tokenization")
-            with token_timer.timing():
-                inputs = tokenizer(
-                    valid_texts,
-                    padding=True,
-                    truncation=True,
-                    max_length=512,
-                    return_tensors="pt"
-                ).to(device)
-
-            # Compute embeddings
-            embed_timer = DeviceTimer("embedding computation")
-            with embed_timer.timing():
-                with torch.no_grad():
-                    outputs = model(**inputs)
-                    hidden_states = outputs.last_hidden_state
-                    
-                    # Mean pooling
-                    attention_mask = inputs['attention_mask']
-                    mask_expanded = attention_mask.unsqueeze(-1).expand(hidden_states.size()).float()
-                    sum_embeddings = torch.sum(hidden_states * mask_expanded, 1)
-                    sum_mask = torch.clamp(mask_expanded.sum(1), min=1e-9)
-                    batch_embeddings = sum_embeddings / sum_mask
-                embed_timer.print_elapsed()
-
-            return batch_embeddings.cpu().numpy()
-
-        # ZMQ server main loop - modified to use REP socket
-        context = zmq.Context()
-        socket = context.socket(zmq.ROUTER)  # Changed to REP socket
-        socket.bind(f"tcp://127.0.0.1:{zmq_port}")
-        print(f"INFO: ZMQ ROUTER server listening on port {zmq_port}")
-
-        # Set timeouts
-        socket.setsockopt(zmq.RCVTIMEO, 5000)  # 5 second receive timeout
-        socket.setsockopt(zmq.SNDTIMEO, 300000)  # 300 second send timeout
-
-        from . import embedding_pb2
-
-        print(f"INFO: Embedding server ready to serve requests")
-
-        # Start warmup thread if enabled
-        if enable_warmup and len(passages) > 0:
-            import threading
-            print(f"Warmup enabled: starting warmup thread")
-            warmup_thread = threading.Thread(target=client_warmup, args=(zmq_port,))
-            warmup_thread.daemon = True
-            warmup_thread.start()
-        else:
-            print(f"Warmup disabled or no passages available (enable_warmup={enable_warmup}, passages={len(passages)})")
-
-        while True:
-            try:
-                parts = socket.recv_multipart()
-
-                # --- Restore robust message format detection ---
-                # Must check parts length to avoid IndexError
-                if len(parts) >= 3:
-                    identity = parts[0]
-                    # empty = parts[1]  # We usually don't care about the middle empty frame
-                    message = parts[2]
-                elif len(parts) == 2:
-                    # Can also handle cases without empty frame
-                    identity = parts[0]
-                    message = parts[1]
-                else:
-                    # If received message format is wrong, print warning and ignore it instead of crashing
-                    print(f"WARNING: Received unexpected message format with {len(parts)} parts. Ignoring.")
-                    continue
-                print(f"INFO: Received ZMQ request from client {identity.hex()[:8]}, size {len(message)} bytes")
-
-                # Handle control messages (MessagePack format)
-                try:
-                    request_payload = msgpack.unpackb(message)
-                    if isinstance(request_payload, list) and len(request_payload) >= 1:
-                        if request_payload[0] == "__QUERY_META_PATH__":
-                            # Return the current meta path being used by the server
-                            current_meta_path = getattr(passages, '_meta_path', '') if hasattr(passages, '_meta_path') else ''
-                            response = [current_meta_path]
-                            socket.send_multipart([identity, b'', msgpack.packb(response)])
-                            continue
-                            
-                        elif request_payload[0] == "__UPDATE_META_PATH__" and len(request_payload) >= 2:
-                            # Update the server's meta path and reload passages
-                            new_meta_path = request_payload[1]
-                            try:
-                                print(f"INFO: Updating server meta path to: {new_meta_path}")
-                                # Reload passages from the new meta file
-                                passages = load_passages_from_metadata(new_meta_path)
-                                # Store the meta path for future queries
-                                passages._meta_path = new_meta_path
-                                response = ["SUCCESS"]
-                                print(f"INFO: Successfully updated meta path and reloaded {len(passages)} passages")
-                            except Exception as e:
-                                print(f"ERROR: Failed to update meta path: {e}")
-                                response = ["FAILED", str(e)]
-                            socket.send_multipart([identity, b'', msgpack.packb(response)])
-                            continue
-                            
-                        elif request_payload[0] == "__QUERY_MODEL__":
-                            # Return the current model being used by the server
-                            response = [model_name]
-                            socket.send_multipart([identity, b'', msgpack.packb(response)])
-                            continue
-                            
-                        elif request_payload[0] == "__UPDATE_MODEL__" and len(request_payload) >= 2:
-                            # Update the server's embedding model
-                            new_model_name = request_payload[1]
-                            try:
-                                print(f"INFO: Updating server model from {model_name} to: {new_model_name}")
-                                
-                                # Clean up old model to free memory
-                                if not use_mlx:
-                                    print("INFO: Releasing old model from memory...")
-                                    old_model = model
-                                    old_tokenizer = tokenizer
-                                    
-                                    # Load new tokenizer first
-                                    print(f"Loading new tokenizer for {new_model_name}...")
-                                    tokenizer = AutoTokenizer.from_pretrained(new_model_name, use_fast=True)
-                                    
-                                    # Load new model
-                                    print(f"Loading new model {new_model_name}...")
-                                    model = AutoModel.from_pretrained(new_model_name).to(device).eval()
-                                    
-                                    # Optimize new model
-                                    if cuda_available or mps_available:
-                                        try:
-                                            model = model.half()
-                                            model = torch.compile(model)
-                                            print(f"INFO: Using FP16 precision with model: {new_model_name}")
-                                        except Exception as e:
-                                            print(f"WARNING: Model optimization failed: {e}")
-                                    
-                                    # Now safely delete old model after new one is loaded
-                                    del old_model
-                                    del old_tokenizer
-                                    
-                                    # Clear GPU cache if available
-                                    if device.type == "cuda":
-                                        torch.cuda.empty_cache()
-                                        print("INFO: Cleared CUDA cache")
-                                    elif device.type == "mps":
-                                        torch.mps.empty_cache()
-                                        print("INFO: Cleared MPS cache")
-                                    
-                                    # Force garbage collection
-                                    import gc
-                                    gc.collect()
-                                    print("INFO: Memory cleanup completed")
-                                
-                                # Update model name
-                                model_name = new_model_name
-                                
-                                response = ["SUCCESS"]
-                                print(f"INFO: Successfully updated model to: {new_model_name}")
-                            except Exception as e:
-                                print(f"ERROR: Failed to update model: {e}")
-                                response = ["FAILED", str(e)]
-                            socket.send_multipart([identity, b'', msgpack.packb(response)])
-                            continue
-                except:
-                    # Not a control message, continue with normal protobuf processing
-                    pass
-
-                e2e_start = time.time()
-                lookup_timer = DeviceTimer("text lookup")
-
-                # Parse request
-                req_proto = embedding_pb2.NodeEmbeddingRequest()
-                req_proto.ParseFromString(message)
-                node_ids = req_proto.node_ids
-                print(f"INFO: Request for {len(node_ids)} node embeddings: {list(node_ids)}")
-
-                # Add debug information
-                if len(node_ids) > 0:
-                    print(f"DEBUG: Node ID range: {min(node_ids)} to {max(node_ids)}")
-                
-                # Look up texts
-                texts = []
-                missing_ids = []
-                with lookup_timer.timing():
-                    for nid in node_ids:
-                        txtinfo = passages[nid]
-                        txt = txtinfo["text"]
-                        if txt:
-                            texts.append(txt)
-                        else:
-                            # If text is empty, we still need a placeholder for batch processing,
-                            # but record its ID as missing
-                            texts.append("") 
-                            missing_ids.append(nid)
-                lookup_timer.print_elapsed()
-
-                if missing_ids:
-                    print(f"WARNING: Missing passages for IDs: {missing_ids}")
-
-                # Process batch
-                total_size = len(texts)
-                print(f"INFO: Total batch size: {total_size}, max_batch_size: {max_batch_size}")
-                
-                all_embeddings = []
-                
-                if total_size > max_batch_size:
-                    print(f"INFO: Splitting batch of size {total_size} into chunks of {max_batch_size}")
-                    for i in range(0, total_size, max_batch_size):
-                        end_idx = min(i + max_batch_size, total_size)
-                        print(f"INFO: Processing chunk {i//max_batch_size + 1}/{(total_size + max_batch_size - 1)//max_batch_size}: items {i} to {end_idx-1}")
-                        
-                        chunk_texts = texts[i:end_idx]
-                        chunk_ids = node_ids[i:end_idx]
-                        
-                        if embedding_mode == "mlx":
-                            embeddings_chunk = compute_embeddings_mlx(chunk_texts, model_name, batch_size=16)
-                        elif embedding_mode == "openai":
-                            embeddings_chunk = compute_embeddings_openai(chunk_texts, model_name)
-                        else:  # sentence-transformers
-                            embeddings_chunk = process_batch_pytorch(chunk_texts, chunk_ids, missing_ids)
-                        all_embeddings.append(embeddings_chunk)
-                        
-                        if embedding_mode == "sentence-transformers":
-                            if cuda_available:
-                                torch.cuda.empty_cache()
-                            elif device.type == "mps":
-                                torch.mps.empty_cache()
-                            
-                    hidden = np.vstack(all_embeddings)
-                    print(f"INFO: Combined embeddings shape: {hidden.shape}")
-                else:
-                    if embedding_mode == "mlx":
-                        hidden = compute_embeddings_mlx(texts, model_name, batch_size=16)
-                    elif embedding_mode == "openai":
-                        hidden = compute_embeddings_openai(texts, model_name)
-                    else:  # sentence-transformers
-                        hidden = process_batch_pytorch(texts, node_ids, missing_ids)
-
-                # Serialize response
-                ser_start = time.time()
-
-                resp_proto = embedding_pb2.NodeEmbeddingResponse()
-                hidden_contiguous = np.ascontiguousarray(hidden, dtype=np.float32)
-                resp_proto.embeddings_data = hidden_contiguous.tobytes()
-                resp_proto.dimensions.append(hidden_contiguous.shape[0])
-                resp_proto.dimensions.append(hidden_contiguous.shape[1])
-                resp_proto.missing_ids.extend(missing_ids)
-
-                response_data = resp_proto.SerializeToString()
-                
-                # REP socket sends a single response
-                socket.send_multipart([identity, b'', response_data])
-
-                ser_end = time.time()
-
-                print(f"INFO: Serialize time: {ser_end - ser_start:.6f} seconds")
-
-                if embedding_mode == "sentence-transformers":
-                    if device.type == "cuda":
-                        torch.cuda.synchronize()
-                    elif device.type == "mps":
-                        torch.mps.synchronize()
-                e2e_end = time.time()
-                print(f"INFO: ZMQ E2E time: {e2e_end - e2e_start:.6f} seconds")
-
-            except zmq.Again:
-                print("INFO: ZMQ socket timeout, continuing to listen")
-                continue
-            except Exception as e:
-                print(f"ERROR: Error in ZMQ server: {e}")
-                try:
-                    # Send empty response to maintain REQ-REP state
-                    empty_resp = embedding_pb2.NodeEmbeddingResponse()
-                    socket.send(empty_resp.SerializeToString())
-                except:
-                    # If sending fails, recreate socket
-                    socket.close()
-                    socket = context.socket(zmq.REP)
-                    socket.bind(f"tcp://127.0.0.1:{zmq_port}")
-                    socket.setsockopt(zmq.RCVTIMEO, 5000)
-                    socket.setsockopt(zmq.SNDTIMEO, 300000)
-                    print("INFO: ZMQ socket recreated after error")
-
-    except Exception as e:
-        print(f"ERROR: Failed to start embedding server: {e}")
-        raise
-
-
-def create_embedding_server(
-    domain="demo",
-    load_passages=True,
-    load_embeddings=False,
-    use_fp16=True,
-    use_int8=False,
-    use_cuda_graphs=False,
-    zmq_port=5555,
-    max_batch_size=128,
-    lazy_load_passages=False,
-    model_name="sentence-transformers/all-mpnet-base-v2",
-    passages_file: Optional[str] = None,
-    embedding_mode: str = "sentence-transformers",
-    enable_warmup: bool = False,
-):
-    """
-    原有的 create_embedding_server 函数保持不变
-    这个是阻塞版本，用于直接运行
-    """
-    create_embedding_server_thread(zmq_port, model_name, max_batch_size, passages_file, embedding_mode, enable_warmup)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Embedding service")
-    parser.add_argument("--zmq-port", type=int, default=5555, help="ZMQ port to run on")
-    parser.add_argument("--domain", type=str, default="demo", help="Domain name")
-    parser.add_argument("--passages-file", type=str, help="JSON file containing passage ID to text mapping")
-    parser.add_argument("--load-passages", action="store_true", default=True)
-    parser.add_argument("--load-embeddings", action="store_true", default=False)
-    parser.add_argument("--use-fp16", action="store_true", default=False)
-    parser.add_argument("--use-int8", action="store_true", default=False)
-    parser.add_argument("--use-cuda-graphs", action="store_true", default=False)
-    parser.add_argument("--max-batch-size", type=int, default=128, help="Maximum batch size before splitting")
-    parser.add_argument("--lazy-load-passages", action="store_true", default=True)
-    parser.add_argument("--model-name", type=str, default="sentence-transformers/all-mpnet-base-v2", 
-                        help="Embedding model name")
-    parser.add_argument("--embedding-mode", type=str, default="sentence-transformers", 
-                        choices=["sentence-transformers", "mlx", "openai"],
-                        help="Embedding backend mode")
-    parser.add_argument("--use-mlx", action="store_true", default=False, help="Use MLX backend for embeddings (deprecated: use --embedding-mode mlx)")
-    parser.add_argument("--disable-warmup", action="store_true", default=False, help="Disable warmup requests on server start")
-    args = parser.parse_args()
-    
-    # Handle backward compatibility with use_mlx
-    embedding_mode = args.embedding_mode
-    if args.use_mlx:
-        embedding_mode = "mlx"
-
-    create_embedding_server(
-        domain=args.domain,
-        load_passages=args.load_passages,
-        load_embeddings=args.load_embeddings,
-        use_fp16=args.use_fp16,
-        use_int8=args.use_int8,
-        use_cuda_graphs=args.use_cuda_graphs,
-        zmq_port=args.zmq_port,
-        max_batch_size=args.max_batch_size,
-        lazy_load_passages=args.lazy_load_passages,
-        model_name=args.model_name,
-        passages_file=args.passages_file,
-        embedding_mode=embedding_mode,
-        enable_warmup=not args.disable_warmup,
-    )

packages/leann-backend-diskann/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "leann-backend-diskann"
-version = "0.1.0"
-dependencies = ["leann-core==0.1.0", "numpy"]
+version = "0.1.16"
+dependencies = ["leann-core==0.1.16", "numpy", "protobuf>=3.19.0"]
 
 [tool.scikit-build]
 # Key: simplified CMake path
@@ -16,4 +16,4 @@ wheel.packages = ["leann_backend_diskann"]
 editable.mode = "redirect"
 cmake.build-type = "Release"
 build.verbose = true
-build.tool-args = ["-j8"]
+build.tool-args = ["-j8"]

packages/leann-backend-diskann/third_party/embedding.proto

@@ -2,12 +2,12 @@ syntax = "proto3";
 
 package protoembedding;
 
-message NodeEmbeddingRequest { 
-  repeated uint32 node_ids = 1; 
+message NodeEmbeddingRequest {
+  repeated uint32 node_ids = 1;
 }
 
 message NodeEmbeddingResponse {
   bytes embeddings_data = 1;        // All embedded binary datas
   repeated int32 dimensions = 2;    // Shape [batch_size, embedding_dim]
   repeated uint32 missing_ids = 3;  // Missing node ids
-}
+}

packages/leann-backend-hnsw/CMakeLists.txt

@@ -10,6 +10,14 @@ if(APPLE)
     set(OpenMP_C_LIB_NAMES "omp")
     set(OpenMP_CXX_LIB_NAMES "omp")
     set(OpenMP_omp_LIBRARY "/opt/homebrew/opt/libomp/lib/libomp.dylib")
+
+    # Force use of system libc++ to avoid version mismatch
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -stdlib=libc++")
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -stdlib=libc++")
+    set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -stdlib=libc++")
+
+    # Set minimum macOS version for better compatibility
+    set(CMAKE_OSX_DEPLOYMENT_TARGET "11.0" CACHE STRING "Minimum macOS version")
 endif()
 
 # Use system ZeroMQ instead of building from source
@@ -52,4 +60,4 @@ set(FAISS_BUILD_AVX512 OFF CACHE BOOL "" FORCE)
 # IMPORTANT: Disable building AVX versions to speed up compilation
 set(FAISS_BUILD_AVX_VERSIONS OFF CACHE BOOL "" FORCE)
 
-add_subdirectory(third_party/faiss)
+add_subdirectory(third_party/faiss)

packages/leann-backend-hnsw/leann_backend_hnsw/__init__.py

@@ -1 +1 @@
-from . import hnsw_backend
+from . import hnsw_backend as hnsw_backend

packages/leann-backend-hnsw/leann_backend_hnsw/convert_to_csr.py

@@ -1,87 +1,115 @@
+import argparse
+import gc  # Import garbage collector interface
+import os
 import struct
 import sys
-import numpy as np
-import os
-import argparse
-import gc # Import garbage collector interface
 import time
+
+import numpy as np
+
 # --- FourCCs (add more if needed) ---
-INDEX_HNSW_FLAT_FOURCC = int.from_bytes(b'IHNf', 'little')
+INDEX_HNSW_FLAT_FOURCC = int.from_bytes(b"IHNf", "little")
 # Add other HNSW fourccs if you expect different storage types inside HNSW
 # INDEX_HNSW_PQ_FOURCC = int.from_bytes(b'IHNp', 'little')
 # INDEX_HNSW_SQ_FOURCC = int.from_bytes(b'IHNs', 'little')
 # INDEX_HNSW_CAGRA_FOURCC = int.from_bytes(b'IHNc', 'little') # Example
 
-EXPECTED_HNSW_FOURCCS = {INDEX_HNSW_FLAT_FOURCC} # Modify if needed
-NULL_INDEX_FOURCC = int.from_bytes(b'null', 'little')
+EXPECTED_HNSW_FOURCCS = {INDEX_HNSW_FLAT_FOURCC}  # Modify if needed
+NULL_INDEX_FOURCC = int.from_bytes(b"null", "little")
 
 # --- Helper functions for reading/writing binary data ---
 
+
 def read_struct(f, fmt):
     """Reads data according to the struct format."""
     size = struct.calcsize(fmt)
     data = f.read(size)
     if len(data) != size:
-        raise EOFError(f"File ended unexpectedly reading struct fmt '{fmt}'. Expected {size} bytes, got {len(data)}.")
+        raise EOFError(
+            f"File ended unexpectedly reading struct fmt '{fmt}'. Expected {size} bytes, got {len(data)}."
+        )
     return struct.unpack(fmt, data)[0]
 
+
 def read_vector_raw(f, element_fmt_char):
     """Reads a vector (size followed by data), returns count and raw bytes."""
-    count = -1 # Initialize count
-    total_bytes = -1 # Initialize total_bytes
+    count = -1  # Initialize count
+    total_bytes = -1  # Initialize total_bytes
     try:
-        count = read_struct(f, '<Q') # size_t usually 64-bit unsigned
+        count = read_struct(f, "<Q")  # size_t usually 64-bit unsigned
         element_size = struct.calcsize(element_fmt_char)
         # --- FIX for MemoryError: Check for unreasonably large count ---
-        max_reasonable_count = 10 * (10**9) # ~10 billion elements limit
+        max_reasonable_count = 10 * (10**9)  # ~10 billion elements limit
         if count > max_reasonable_count or count < 0:
-            raise MemoryError(f"Vector count {count} seems unreasonably large, possibly due to file corruption or incorrect format read.")
+            raise MemoryError(
+                f"Vector count {count} seems unreasonably large, possibly due to file corruption or incorrect format read."
+            )
 
         total_bytes = count * element_size
         # --- FIX for MemoryError: Check for huge byte size before allocation ---
-        max_reasonable_bytes = 50 * (1024**3) # ~50 GB limit
-        if total_bytes > max_reasonable_bytes or total_bytes < 0: # Check for overflow
-             raise MemoryError(f"Attempting to read {total_bytes} bytes ({count} elements * {element_size} bytes/element), which exceeds the safety limit. File might be corrupted or format mismatch.")
+        max_reasonable_bytes = 50 * (1024**3)  # ~50 GB limit
+        if total_bytes > max_reasonable_bytes or total_bytes < 0:  # Check for overflow
+            raise MemoryError(
+                f"Attempting to read {total_bytes} bytes ({count} elements * {element_size} bytes/element), which exceeds the safety limit. File might be corrupted or format mismatch."
+            )
 
         data_bytes = f.read(total_bytes)
 
         if len(data_bytes) != total_bytes:
-             raise EOFError(f"File ended unexpectedly reading vector data. Expected {total_bytes} bytes, got {len(data_bytes)}.")
+            raise EOFError(
+                f"File ended unexpectedly reading vector data. Expected {total_bytes} bytes, got {len(data_bytes)}."
+            )
         return count, data_bytes
     except (MemoryError, OverflowError) as e:
-         # Add context to the error message
-         print(f"\nError during raw vector read (element_fmt='{element_fmt_char}', count={count}, total_bytes={total_bytes}): {e}", file=sys.stderr)
-         raise e # Re-raise the original error type
+        # Add context to the error message
+        print(
+            f"\nError during raw vector read (element_fmt='{element_fmt_char}', count={count}, total_bytes={total_bytes}): {e}",
+            file=sys.stderr,
+        )
+        raise e  # Re-raise the original error type
+
 
 def read_numpy_vector(f, np_dtype, struct_fmt_char):
     """Reads a vector into a NumPy array."""
-    count = -1 # Initialize count for robust error handling
-    print(f"  Reading vector (dtype={np_dtype}, fmt='{struct_fmt_char}')... ", end='', flush=True)
+    count = -1  # Initialize count for robust error handling
+    print(
+        f"  Reading vector (dtype={np_dtype}, fmt='{struct_fmt_char}')... ",
+        end="",
+        flush=True,
+    )
     try:
         count, data_bytes = read_vector_raw(f, struct_fmt_char)
         print(f"Count={count}, Bytes={len(data_bytes)}")
         if count > 0 and len(data_bytes) > 0:
             arr = np.frombuffer(data_bytes, dtype=np_dtype)
             if arr.size != count:
-                raise ValueError(f"Inconsistent array size after reading. Expected {count}, got {arr.size}")
+                raise ValueError(
+                    f"Inconsistent array size after reading. Expected {count}, got {arr.size}"
+                )
             return arr
         elif count == 0:
-             return np.array([], dtype=np_dtype)
+            return np.array([], dtype=np_dtype)
         else:
-             raise ValueError("Read zero bytes but count > 0.")
+            raise ValueError("Read zero bytes but count > 0.")
     except MemoryError as e:
         # Now count should be defined (or -1 if error was in read_struct)
-        print(f"\nMemoryError creating NumPy array (dtype={np_dtype}, count={count}). {e}", file=sys.stderr)
+        print(
+            f"\nMemoryError creating NumPy array (dtype={np_dtype}, count={count}). {e}",
+            file=sys.stderr,
+        )
         raise e
-    except Exception as e: # Catch other potential errors like ValueError
-        print(f"\nError reading numpy vector (dtype={np_dtype}, fmt='{struct_fmt_char}', count={count}): {e}", file=sys.stderr)
+    except Exception as e:  # Catch other potential errors like ValueError
+        print(
+            f"\nError reading numpy vector (dtype={np_dtype}, fmt='{struct_fmt_char}', count={count}): {e}",
+            file=sys.stderr,
+        )
         raise e
 
 
 def write_numpy_vector(f, arr, struct_fmt_char):
     """Writes a NumPy array as a vector (size followed by data)."""
     count = arr.size
-    f.write(struct.pack('<Q', count))
+    f.write(struct.pack("<Q", count))
     try:
         expected_dtype = np.dtype(struct_fmt_char)
         if arr.dtype != expected_dtype:
@@ -89,23 +117,30 @@ def write_numpy_vector(f, arr, struct_fmt_char):
         else:
             data_to_write = arr.tobytes()
         f.write(data_to_write)
-        del data_to_write # Hint GC
+        del data_to_write  # Hint GC
     except MemoryError as e:
-         print(f"\nMemoryError converting NumPy array to bytes for writing (size={count}, dtype={arr.dtype}). {e}", file=sys.stderr)
-         raise e
+        print(
+            f"\nMemoryError converting NumPy array to bytes for writing (size={count}, dtype={arr.dtype}). {e}",
+            file=sys.stderr,
+        )
+        raise e
+
 
 def write_list_vector(f, lst, struct_fmt_char):
     """Writes a Python list as a vector iteratively."""
     count = len(lst)
-    f.write(struct.pack('<Q', count))
-    fmt = '<' + struct_fmt_char
+    f.write(struct.pack("<Q", count))
+    fmt = "<" + struct_fmt_char
     chunk_size = 1024 * 1024
     element_size = struct.calcsize(fmt)
     # Allocate buffer outside the loop if possible, or handle MemoryError during allocation
     try:
         buffer = bytearray(chunk_size * element_size)
     except MemoryError:
-        print(f"MemoryError: Cannot allocate buffer for writing list vector chunk (size {chunk_size * element_size} bytes).", file=sys.stderr)
+        print(
+            f"MemoryError: Cannot allocate buffer for writing list vector chunk (size {chunk_size * element_size} bytes).",
+            file=sys.stderr,
+        )
         raise
     buffer_count = 0
 
@@ -116,66 +151,80 @@ def write_list_vector(f, lst, struct_fmt_char):
             buffer_count += 1
 
             if buffer_count == chunk_size or i == count - 1:
-                f.write(buffer[:buffer_count * element_size])
+                f.write(buffer[: buffer_count * element_size])
                 buffer_count = 0
 
         except struct.error as e:
-            print(f"\nStruct packing error for item {item} at index {i} with format '{fmt}'. {e}", file=sys.stderr)
+            print(
+                f"\nStruct packing error for item {item} at index {i} with format '{fmt}'. {e}",
+                file=sys.stderr,
+            )
             raise e
 
 
 def get_cum_neighbors(cum_nneighbor_per_level_np, level):
     """Helper to get cumulative neighbors count, matching C++ logic."""
-    if level < 0: return 0
+    if level < 0:
+        return 0
     if level < len(cum_nneighbor_per_level_np):
         return cum_nneighbor_per_level_np[level]
     else:
         return cum_nneighbor_per_level_np[-1] if len(cum_nneighbor_per_level_np) > 0 else 0
 
-def write_compact_format(f_out, original_hnsw_data, assign_probas_np, cum_nneighbor_per_level_np, 
-                        levels_np, compact_level_ptr, compact_node_offsets_np, 
-                        compact_neighbors_data, storage_fourcc, storage_data):
+
+def write_compact_format(
+    f_out,
+    original_hnsw_data,
+    assign_probas_np,
+    cum_nneighbor_per_level_np,
+    levels_np,
+    compact_level_ptr,
+    compact_node_offsets_np,
+    compact_neighbors_data,
+    storage_fourcc,
+    storage_data,
+):
     """Write HNSW data in compact format following C++ read order exactly."""
     # Write IndexHNSW Header
-    f_out.write(struct.pack('<I', original_hnsw_data['index_fourcc']))
-    f_out.write(struct.pack('<i', original_hnsw_data['d']))
-    f_out.write(struct.pack('<q', original_hnsw_data['ntotal']))
-    f_out.write(struct.pack('<q', original_hnsw_data['dummy1']))
-    f_out.write(struct.pack('<q', original_hnsw_data['dummy2']))
-    f_out.write(struct.pack('<?', original_hnsw_data['is_trained']))
-    f_out.write(struct.pack('<i', original_hnsw_data['metric_type']))
-    if original_hnsw_data['metric_type'] > 1:
-         f_out.write(struct.pack('<f', original_hnsw_data['metric_arg']))
+    f_out.write(struct.pack("<I", original_hnsw_data["index_fourcc"]))
+    f_out.write(struct.pack("<i", original_hnsw_data["d"]))
+    f_out.write(struct.pack("<q", original_hnsw_data["ntotal"]))
+    f_out.write(struct.pack("<q", original_hnsw_data["dummy1"]))
+    f_out.write(struct.pack("<q", original_hnsw_data["dummy2"]))
+    f_out.write(struct.pack("<?", original_hnsw_data["is_trained"]))
+    f_out.write(struct.pack("<i", original_hnsw_data["metric_type"]))
+    if original_hnsw_data["metric_type"] > 1:
+        f_out.write(struct.pack("<f", original_hnsw_data["metric_arg"]))
 
     # Write HNSW struct parts (standard order)
-    write_numpy_vector(f_out, assign_probas_np, 'd')
-    write_numpy_vector(f_out, cum_nneighbor_per_level_np, 'i')
-    write_numpy_vector(f_out, levels_np, 'i')
+    write_numpy_vector(f_out, assign_probas_np, "d")
+    write_numpy_vector(f_out, cum_nneighbor_per_level_np, "i")
+    write_numpy_vector(f_out, levels_np, "i")
 
     # Write compact format flag
-    f_out.write(struct.pack('<?', True)) # storage_is_compact = True
+    f_out.write(struct.pack("<?", True))  # storage_is_compact = True
 
     # Write compact data in CORRECT C++ read order: level_ptr, node_offsets FIRST
     if isinstance(compact_level_ptr, np.ndarray):
-        write_numpy_vector(f_out, compact_level_ptr, 'Q')
+        write_numpy_vector(f_out, compact_level_ptr, "Q")
     else:
-        write_list_vector(f_out, compact_level_ptr, 'Q')
-    
-    write_numpy_vector(f_out, compact_node_offsets_np, 'Q')
+        write_list_vector(f_out, compact_level_ptr, "Q")
+
+    write_numpy_vector(f_out, compact_node_offsets_np, "Q")
 
     # Write HNSW scalar parameters
-    f_out.write(struct.pack('<i', original_hnsw_data['entry_point']))
-    f_out.write(struct.pack('<i', original_hnsw_data['max_level']))
-    f_out.write(struct.pack('<i', original_hnsw_data['efConstruction']))
-    f_out.write(struct.pack('<i', original_hnsw_data['efSearch']))
-    f_out.write(struct.pack('<i', original_hnsw_data['dummy_upper_beam']))
+    f_out.write(struct.pack("<i", original_hnsw_data["entry_point"]))
+    f_out.write(struct.pack("<i", original_hnsw_data["max_level"]))
+    f_out.write(struct.pack("<i", original_hnsw_data["efConstruction"]))
+    f_out.write(struct.pack("<i", original_hnsw_data["efSearch"]))
+    f_out.write(struct.pack("<i", original_hnsw_data["dummy_upper_beam"]))
 
     # Write storage fourcc (this determines how to read what follows)
-    f_out.write(struct.pack('<I', storage_fourcc))
-    
+    f_out.write(struct.pack("<I", storage_fourcc))
+
     # Write compact neighbors data AFTER storage fourcc
-    write_list_vector(f_out, compact_neighbors_data, 'i')
-    
+    write_list_vector(f_out, compact_neighbors_data, "i")
+
     # Write storage data if not NULL (only after neighbors)
     if storage_fourcc != NULL_INDEX_FOURCC and storage_data:
         f_out.write(storage_data)
@@ -183,11 +232,12 @@ def write_compact_format(f_out, original_hnsw_data, assign_probas_np, cum_nneigh
 
 # --- Main Conversion Logic ---
 
+
 def convert_hnsw_graph_to_csr(input_filename, output_filename, prune_embeddings=True):
     """
     Converts an HNSW graph file to the CSR format.
     Supports both original and already-compact formats (backward compatibility).
-    
+
     Args:
         input_filename: Input HNSW index file
         output_filename: Output CSR index file
@@ -196,172 +246,228 @@ def convert_hnsw_graph_to_csr(input_filename, output_filename, prune_embeddings=
     print(f"Starting conversion: {input_filename} -> {output_filename}")
     start_time = time.time()
     original_hnsw_data = {}
-    neighbors_np = None # Initialize to allow check in finally block
+    neighbors_np = None  # Initialize to allow check in finally block
     try:
-        with open(input_filename, 'rb') as f_in, open(output_filename, 'wb') as f_out:
-
+        with open(input_filename, "rb") as f_in, open(output_filename, "wb") as f_out:
             # --- Read IndexHNSW FourCC and Header ---
             print(f"[{time.time() - start_time:.2f}s] Reading Index HNSW header...")
             # ... (Keep the header reading logic as before) ...
-            hnsw_index_fourcc = read_struct(f_in, '<I')
+            hnsw_index_fourcc = read_struct(f_in, "<I")
             if hnsw_index_fourcc not in EXPECTED_HNSW_FOURCCS:
-                 print(f"Error: Expected HNSW Index FourCC ({list(EXPECTED_HNSW_FOURCCS)}), got {hnsw_index_fourcc:08x}.", file=sys.stderr)
-                 return False
-            original_hnsw_data['index_fourcc'] = hnsw_index_fourcc
-            original_hnsw_data['d'] = read_struct(f_in, '<i')
-            original_hnsw_data['ntotal'] = read_struct(f_in, '<q')
-            original_hnsw_data['dummy1'] = read_struct(f_in, '<q')
-            original_hnsw_data['dummy2'] = read_struct(f_in, '<q')
-            original_hnsw_data['is_trained'] = read_struct(f_in, '?')
-            original_hnsw_data['metric_type'] = read_struct(f_in, '<i')
-            original_hnsw_data['metric_arg'] = 0.0
-            if original_hnsw_data['metric_type'] > 1:
-                 original_hnsw_data['metric_arg'] = read_struct(f_in, '<f')
-            print(f"[{time.time() - start_time:.2f}s]   Header read: d={original_hnsw_data['d']}, ntotal={original_hnsw_data['ntotal']}")
-
+                print(
+                    f"Error: Expected HNSW Index FourCC ({list(EXPECTED_HNSW_FOURCCS)}), got {hnsw_index_fourcc:08x}.",
+                    file=sys.stderr,
+                )
+                return False
+            original_hnsw_data["index_fourcc"] = hnsw_index_fourcc
+            original_hnsw_data["d"] = read_struct(f_in, "<i")
+            original_hnsw_data["ntotal"] = read_struct(f_in, "<q")
+            original_hnsw_data["dummy1"] = read_struct(f_in, "<q")
+            original_hnsw_data["dummy2"] = read_struct(f_in, "<q")
+            original_hnsw_data["is_trained"] = read_struct(f_in, "?")
+            original_hnsw_data["metric_type"] = read_struct(f_in, "<i")
+            original_hnsw_data["metric_arg"] = 0.0
+            if original_hnsw_data["metric_type"] > 1:
+                original_hnsw_data["metric_arg"] = read_struct(f_in, "<f")
+            print(
+                f"[{time.time() - start_time:.2f}s]   Header read: d={original_hnsw_data['d']}, ntotal={original_hnsw_data['ntotal']}"
+            )
 
             # --- Read original HNSW struct data ---
             print(f"[{time.time() - start_time:.2f}s] Reading HNSW struct vectors...")
-            assign_probas_np = read_numpy_vector(f_in, np.float64, 'd')
-            print(f"[{time.time() - start_time:.2f}s]   Read assign_probas ({assign_probas_np.size})")
+            assign_probas_np = read_numpy_vector(f_in, np.float64, "d")
+            print(
+                f"[{time.time() - start_time:.2f}s]   Read assign_probas ({assign_probas_np.size})"
+            )
             gc.collect()
 
-            cum_nneighbor_per_level_np = read_numpy_vector(f_in, np.int32, 'i')
-            print(f"[{time.time() - start_time:.2f}s]   Read cum_nneighbor_per_level ({cum_nneighbor_per_level_np.size})")
+            cum_nneighbor_per_level_np = read_numpy_vector(f_in, np.int32, "i")
+            print(
+                f"[{time.time() - start_time:.2f}s]   Read cum_nneighbor_per_level ({cum_nneighbor_per_level_np.size})"
+            )
             gc.collect()
 
-            levels_np = read_numpy_vector(f_in, np.int32, 'i')
+            levels_np = read_numpy_vector(f_in, np.int32, "i")
             print(f"[{time.time() - start_time:.2f}s]   Read levels ({levels_np.size})")
             gc.collect()
 
             ntotal = len(levels_np)
-            if ntotal != original_hnsw_data['ntotal']:
-                 print(f"Warning: ntotal mismatch! Header says {original_hnsw_data['ntotal']}, levels vector size is {ntotal}. Using levels vector size.", file=sys.stderr)
-                 original_hnsw_data['ntotal'] = ntotal
+            if ntotal != original_hnsw_data["ntotal"]:
+                print(
+                    f"Warning: ntotal mismatch! Header says {original_hnsw_data['ntotal']}, levels vector size is {ntotal}. Using levels vector size.",
+                    file=sys.stderr,
+                )
+                original_hnsw_data["ntotal"] = ntotal
 
             # --- Check for compact format flag ---
             print(f"[{time.time() - start_time:.2f}s]   Probing for compact storage flag...")
             pos_before_compact = f_in.tell()
             try:
-                is_compact_flag = read_struct(f_in, '<?')
+                is_compact_flag = read_struct(f_in, "<?")
                 print(f"[{time.time() - start_time:.2f}s]   Found compact flag: {is_compact_flag}")
-                
+
                 if is_compact_flag:
                     # Input is already in compact format - read compact data
-                    print(f"[{time.time() - start_time:.2f}s]   Input is already in compact format, reading compact data...")
-                    
-                    compact_level_ptr = read_numpy_vector(f_in, np.uint64, 'Q')
-                    print(f"[{time.time() - start_time:.2f}s]   Read compact_level_ptr ({compact_level_ptr.size})")
-                    
-                    compact_node_offsets_np = read_numpy_vector(f_in, np.uint64, 'Q')
-                    print(f"[{time.time() - start_time:.2f}s]   Read compact_node_offsets ({compact_node_offsets_np.size})")
-                    
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Input is already in compact format, reading compact data..."
+                    )
+
+                    compact_level_ptr = read_numpy_vector(f_in, np.uint64, "Q")
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Read compact_level_ptr ({compact_level_ptr.size})"
+                    )
+
+                    compact_node_offsets_np = read_numpy_vector(f_in, np.uint64, "Q")
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Read compact_node_offsets ({compact_node_offsets_np.size})"
+                    )
+
                     # Read scalar parameters
-                    original_hnsw_data['entry_point'] = read_struct(f_in, '<i')
-                    original_hnsw_data['max_level'] = read_struct(f_in, '<i')
-                    original_hnsw_data['efConstruction'] = read_struct(f_in, '<i')
-                    original_hnsw_data['efSearch'] = read_struct(f_in, '<i')
-                    original_hnsw_data['dummy_upper_beam'] = read_struct(f_in, '<i')
-                    print(f"[{time.time() - start_time:.2f}s]   Read scalar params (ep={original_hnsw_data['entry_point']}, max_lvl={original_hnsw_data['max_level']})")
+                    original_hnsw_data["entry_point"] = read_struct(f_in, "<i")
+                    original_hnsw_data["max_level"] = read_struct(f_in, "<i")
+                    original_hnsw_data["efConstruction"] = read_struct(f_in, "<i")
+                    original_hnsw_data["efSearch"] = read_struct(f_in, "<i")
+                    original_hnsw_data["dummy_upper_beam"] = read_struct(f_in, "<i")
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Read scalar params (ep={original_hnsw_data['entry_point']}, max_lvl={original_hnsw_data['max_level']})"
+                    )
 
                     # Read storage fourcc
-                    storage_fourcc = read_struct(f_in, '<I')
-                    print(f"[{time.time() - start_time:.2f}s]   Found storage fourcc: {storage_fourcc:08x}")
-                    
+                    storage_fourcc = read_struct(f_in, "<I")
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Found storage fourcc: {storage_fourcc:08x}"
+                    )
+
                     if prune_embeddings and storage_fourcc != NULL_INDEX_FOURCC:
                         # Read compact neighbors data
-                        compact_neighbors_data_np = read_numpy_vector(f_in, np.int32, 'i')
-                        print(f"[{time.time() - start_time:.2f}s]   Read compact neighbors data ({compact_neighbors_data_np.size})")
+                        compact_neighbors_data_np = read_numpy_vector(f_in, np.int32, "i")
+                        print(
+                            f"[{time.time() - start_time:.2f}s]   Read compact neighbors data ({compact_neighbors_data_np.size})"
+                        )
                         compact_neighbors_data = compact_neighbors_data_np.tolist()
                         del compact_neighbors_data_np
-                        
+
                         # Skip storage data and write with NULL marker
-                        print(f"[{time.time() - start_time:.2f}s]   Pruning embeddings: Writing NULL storage marker.")
+                        print(
+                            f"[{time.time() - start_time:.2f}s]   Pruning embeddings: Writing NULL storage marker."
+                        )
                         storage_fourcc = NULL_INDEX_FOURCC
                     elif not prune_embeddings:
                         # Read and preserve compact neighbors and storage
-                        compact_neighbors_data_np = read_numpy_vector(f_in, np.int32, 'i')
+                        compact_neighbors_data_np = read_numpy_vector(f_in, np.int32, "i")
                         compact_neighbors_data = compact_neighbors_data_np.tolist()
                         del compact_neighbors_data_np
-                        
+
                         # Read remaining storage data
                         storage_data = f_in.read()
                     else:
                         # Already pruned (NULL storage)
-                        compact_neighbors_data_np = read_numpy_vector(f_in, np.int32, 'i')
+                        compact_neighbors_data_np = read_numpy_vector(f_in, np.int32, "i")
                         compact_neighbors_data = compact_neighbors_data_np.tolist()
                         del compact_neighbors_data_np
-                        storage_data = b''
-                    
+                        storage_data = b""
+
                     # Write the updated compact format
                     print(f"[{time.time() - start_time:.2f}s] Writing updated compact format...")
-                    write_compact_format(f_out, original_hnsw_data, assign_probas_np, cum_nneighbor_per_level_np, 
-                                       levels_np, compact_level_ptr, compact_node_offsets_np, 
-                                       compact_neighbors_data, storage_fourcc, storage_data if not prune_embeddings else b'')
-                    
+                    write_compact_format(
+                        f_out,
+                        original_hnsw_data,
+                        assign_probas_np,
+                        cum_nneighbor_per_level_np,
+                        levels_np,
+                        compact_level_ptr,
+                        compact_node_offsets_np,
+                        compact_neighbors_data,
+                        storage_fourcc,
+                        storage_data if not prune_embeddings else b"",
+                    )
+
                     print(f"[{time.time() - start_time:.2f}s] Conversion complete.")
                     return True
-                    
+
                 else:
                     # is_compact=False, rewind and read original format
                     f_in.seek(pos_before_compact)
-                    print(f"[{time.time() - start_time:.2f}s]   Compact flag is False, reading original format...")
-                    
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Compact flag is False, reading original format..."
+                    )
+
             except EOFError:
                 # No compact flag found, assume original format
                 f_in.seek(pos_before_compact)
-                print(f"[{time.time() - start_time:.2f}s]   No compact flag found, assuming original format...")
+                print(
+                    f"[{time.time() - start_time:.2f}s]   No compact flag found, assuming original format..."
+                )
 
             # --- Handle potential extra byte in original format (like C++ code) ---
-            print(f"[{time.time() - start_time:.2f}s]   Probing for potential extra byte before non-compact offsets...")
+            print(
+                f"[{time.time() - start_time:.2f}s]   Probing for potential extra byte before non-compact offsets..."
+            )
             pos_before_probe = f_in.tell()
             try:
-                suspected_flag = read_struct(f_in, '<B')  # Read 1 byte
+                suspected_flag = read_struct(f_in, "<B")  # Read 1 byte
                 if suspected_flag == 0x00:
-                    print(f"[{time.time() - start_time:.2f}s]   Found and consumed an unexpected 0x00 byte.")
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Found and consumed an unexpected 0x00 byte."
+                    )
                 elif suspected_flag == 0x01:
-                    print(f"[{time.time() - start_time:.2f}s]   ERROR: Found 0x01 but is_compact should be False")
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   ERROR: Found 0x01 but is_compact should be False"
+                    )
                     raise ValueError("Inconsistent compact flag state")
                 else:
                     # Rewind - this byte is part of offsets data
                     f_in.seek(pos_before_probe)
-                    print(f"[{time.time() - start_time:.2f}s]   Rewound to original position (byte was 0x{suspected_flag:02x})")
+                    print(
+                        f"[{time.time() - start_time:.2f}s]   Rewound to original position (byte was 0x{suspected_flag:02x})"
+                    )
             except EOFError:
                 f_in.seek(pos_before_probe)
-                print(f"[{time.time() - start_time:.2f}s]   No extra byte found (EOF), proceeding with offsets read")
+                print(
+                    f"[{time.time() - start_time:.2f}s]   No extra byte found (EOF), proceeding with offsets read"
+                )
 
             # --- Read original format data ---
-            offsets_np = read_numpy_vector(f_in, np.uint64, 'Q')
+            offsets_np = read_numpy_vector(f_in, np.uint64, "Q")
             print(f"[{time.time() - start_time:.2f}s]   Read offsets ({offsets_np.size})")
             if len(offsets_np) != ntotal + 1:
-                 raise ValueError(f"Inconsistent offsets size: len(levels)={ntotal} but len(offsets)={len(offsets_np)}")
+                raise ValueError(
+                    f"Inconsistent offsets size: len(levels)={ntotal} but len(offsets)={len(offsets_np)}"
+                )
             gc.collect()
 
             print(f"[{time.time() - start_time:.2f}s]   Attempting to read neighbors vector...")
-            neighbors_np = read_numpy_vector(f_in, np.int32, 'i')
+            neighbors_np = read_numpy_vector(f_in, np.int32, "i")
             print(f"[{time.time() - start_time:.2f}s]   Read neighbors ({neighbors_np.size})")
             expected_neighbors_size = offsets_np[-1] if ntotal > 0 else 0
             if neighbors_np.size != expected_neighbors_size:
-                 print(f"Warning: neighbors vector size mismatch. Expected {expected_neighbors_size} based on offsets, got {neighbors_np.size}.")
+                print(
+                    f"Warning: neighbors vector size mismatch. Expected {expected_neighbors_size} based on offsets, got {neighbors_np.size}."
+                )
             gc.collect()
 
-            original_hnsw_data['entry_point'] = read_struct(f_in, '<i')
-            original_hnsw_data['max_level'] = read_struct(f_in, '<i')
-            original_hnsw_data['efConstruction'] = read_struct(f_in, '<i')
-            original_hnsw_data['efSearch'] = read_struct(f_in, '<i')
-            original_hnsw_data['dummy_upper_beam'] = read_struct(f_in, '<i')
-            print(f"[{time.time() - start_time:.2f}s]   Read scalar params (ep={original_hnsw_data['entry_point']}, max_lvl={original_hnsw_data['max_level']})")
+            original_hnsw_data["entry_point"] = read_struct(f_in, "<i")
+            original_hnsw_data["max_level"] = read_struct(f_in, "<i")
+            original_hnsw_data["efConstruction"] = read_struct(f_in, "<i")
+            original_hnsw_data["efSearch"] = read_struct(f_in, "<i")
+            original_hnsw_data["dummy_upper_beam"] = read_struct(f_in, "<i")
+            print(
+                f"[{time.time() - start_time:.2f}s]   Read scalar params (ep={original_hnsw_data['entry_point']}, max_lvl={original_hnsw_data['max_level']})"
+            )
 
             print(f"[{time.time() - start_time:.2f}s] Checking for storage data...")
             storage_fourcc = None
             try:
-                storage_fourcc = read_struct(f_in, '<I')
-                print(f"[{time.time() - start_time:.2f}s]   Found storage fourcc: {storage_fourcc:08x}.")
+                storage_fourcc = read_struct(f_in, "<I")
+                print(
+                    f"[{time.time() - start_time:.2f}s]   Found storage fourcc: {storage_fourcc:08x}."
+                )
             except EOFError:
-                 print(f"[{time.time() - start_time:.2f}s]   No storage data found (EOF).")
+                print(f"[{time.time() - start_time:.2f}s]   No storage data found (EOF).")
             except Exception as e:
-                 print(f"[{time.time() - start_time:.2f}s]   Error reading potential storage data: {e}")
-
+                print(
+                    f"[{time.time() - start_time:.2f}s]   Error reading potential storage data: {e}"
+                )
 
             # --- Perform Conversion ---
             print(f"[{time.time() - start_time:.2f}s] Converting to CSR format...")
@@ -373,17 +479,21 @@ def convert_hnsw_graph_to_csr(input_filename, output_filename, prune_embeddings=
 
             current_level_ptr_idx = 0
             current_data_idx = 0
-            total_valid_neighbors_counted = 0 # For validation
+            total_valid_neighbors_counted = 0  # For validation
 
             # Optimize calculation by getting slices once per node if possible
             for i in range(ntotal):
-                if i > 0 and i % (ntotal // 100 or 1) == 0: # Log progress roughly every 1%
+                if i > 0 and i % (ntotal // 100 or 1) == 0:  # Log progress roughly every 1%
                     progress = (i / ntotal) * 100
                     elapsed = time.time() - start_time
-                    print(f"\r[{elapsed:.2f}s]   Converting node {i}/{ntotal} ({progress:.1f}%)...", end="")
+                    print(
+                        f"\r[{elapsed:.2f}s]   Converting node {i}/{ntotal} ({progress:.1f}%)...",
+                        end="",
+                    )
 
                 node_max_level = levels_np[i] - 1
-                if node_max_level < -1: node_max_level = -1
+                if node_max_level < -1:
+                    node_max_level = -1
 
                 node_ptr_start_index = current_level_ptr_idx
                 compact_node_offsets_np[i] = node_ptr_start_index
@@ -394,13 +504,17 @@ def convert_hnsw_graph_to_csr(input_filename, output_filename, prune_embeddings=
                 for level in range(node_max_level + 1):
                     compact_level_ptr.append(current_data_idx)
 
-                    begin_orig_np = original_offset_start + get_cum_neighbors(cum_nneighbor_per_level_np, level)
-                    end_orig_np = original_offset_start + get_cum_neighbors(cum_nneighbor_per_level_np, level + 1)
+                    begin_orig_np = original_offset_start + get_cum_neighbors(
+                        cum_nneighbor_per_level_np, level
+                    )
+                    end_orig_np = original_offset_start + get_cum_neighbors(
+                        cum_nneighbor_per_level_np, level + 1
+                    )
 
                     begin_orig = int(begin_orig_np)
                     end_orig = int(end_orig_np)
 
-                    neighbors_len = len(neighbors_np) # Cache length
+                    neighbors_len = len(neighbors_np)  # Cache length
                     begin_orig = min(max(0, begin_orig), neighbors_len)
                     end_orig = min(max(begin_orig, end_orig), neighbors_len)
 
@@ -413,83 +527,117 @@ def convert_hnsw_graph_to_csr(input_filename, output_filename, prune_embeddings=
 
                         if num_valid > 0:
                             # Append valid neighbors
-                            compact_neighbors_data.extend(level_neighbors_slice[valid_neighbors_mask])
+                            compact_neighbors_data.extend(
+                                level_neighbors_slice[valid_neighbors_mask]
+                            )
                             current_data_idx += num_valid
                             total_valid_neighbors_counted += num_valid
 
-
                 compact_level_ptr.append(current_data_idx)
                 current_level_ptr_idx += num_pointers_expected
 
             compact_node_offsets_np[ntotal] = current_level_ptr_idx
-            print(f"\r[{time.time() - start_time:.2f}s]   Conversion loop finished.                        ") # Clear progress line
+            print(
+                f"\r[{time.time() - start_time:.2f}s]   Conversion loop finished.                        "
+            )  # Clear progress line
 
             # --- Validation Checks ---
             print(f"[{time.time() - start_time:.2f}s] Running validation checks...")
             valid_check_passed = True
             # Check 1: Total valid neighbors count
-            print(f"    Checking total valid neighbor count...")
+            print("    Checking total valid neighbor count...")
             expected_valid_count = np.sum(neighbors_np >= 0)
             if total_valid_neighbors_counted != len(compact_neighbors_data):
-                 print(f"Error: Mismatch between counted valid neighbors ({total_valid_neighbors_counted}) and final compact_data size ({len(compact_neighbors_data)})!", file=sys.stderr)
-                 valid_check_passed = False
+                print(
+                    f"Error: Mismatch between counted valid neighbors ({total_valid_neighbors_counted}) and final compact_data size ({len(compact_neighbors_data)})!",
+                    file=sys.stderr,
+                )
+                valid_check_passed = False
             if expected_valid_count != len(compact_neighbors_data):
-                 print(f"Error: Mismatch between NumPy count of valid neighbors ({expected_valid_count}) and final compact_data size ({len(compact_neighbors_data)})!", file=sys.stderr)
-                 valid_check_passed = False
+                print(
+                    f"Error: Mismatch between NumPy count of valid neighbors ({expected_valid_count}) and final compact_data size ({len(compact_neighbors_data)})!",
+                    file=sys.stderr,
+                )
+                valid_check_passed = False
             else:
-                 print(f"    OK: Total valid neighbors = {len(compact_neighbors_data)}")
+                print(f"    OK: Total valid neighbors = {len(compact_neighbors_data)}")
 
             # Check 2: Final pointer indices consistency
-            print(f"    Checking final pointer indices...")
+            print("    Checking final pointer indices...")
             if compact_node_offsets_np[ntotal] != len(compact_level_ptr):
-                 print(f"Error: Final node offset ({compact_node_offsets_np[ntotal]}) doesn't match level_ptr size ({len(compact_level_ptr)})!", file=sys.stderr)
-                 valid_check_passed = False
-            if (len(compact_level_ptr) > 0 and compact_level_ptr[-1] != len(compact_neighbors_data)) or \
-               (len(compact_level_ptr) == 0 and len(compact_neighbors_data) != 0):
-                 last_ptr = compact_level_ptr[-1] if len(compact_level_ptr) > 0 else -1
-                 print(f"Error: Last level pointer ({last_ptr}) doesn't match compact_data size ({len(compact_neighbors_data)})!", file=sys.stderr)
-                 valid_check_passed = False
+                print(
+                    f"Error: Final node offset ({compact_node_offsets_np[ntotal]}) doesn't match level_ptr size ({len(compact_level_ptr)})!",
+                    file=sys.stderr,
+                )
+                valid_check_passed = False
+            if (
+                len(compact_level_ptr) > 0 and compact_level_ptr[-1] != len(compact_neighbors_data)
+            ) or (len(compact_level_ptr) == 0 and len(compact_neighbors_data) != 0):
+                last_ptr = compact_level_ptr[-1] if len(compact_level_ptr) > 0 else -1
+                print(
+                    f"Error: Last level pointer ({last_ptr}) doesn't match compact_data size ({len(compact_neighbors_data)})!",
+                    file=sys.stderr,
+                )
+                valid_check_passed = False
             else:
-                 print(f"    OK: Final pointers match data size.")
+                print("    OK: Final pointers match data size.")
 
             if not valid_check_passed:
-                print("Error: Validation checks failed. Output file might be incorrect.", file=sys.stderr)
+                print(
+                    "Error: Validation checks failed. Output file might be incorrect.",
+                    file=sys.stderr,
+                )
                 # Optional: Exit here if validation fails
                 # return False
 
             # --- Explicitly delete large intermediate arrays ---
-            print(f"[{time.time() - start_time:.2f}s] Deleting original neighbors and offsets arrays...")
+            print(
+                f"[{time.time() - start_time:.2f}s] Deleting original neighbors and offsets arrays..."
+            )
             del neighbors_np
             del offsets_np
             gc.collect()
 
-            print(f"    CSR Stats: |data|={len(compact_neighbors_data)}, |level_ptr|={len(compact_level_ptr)}")
+            print(
+                f"    CSR Stats: |data|={len(compact_neighbors_data)}, |level_ptr|={len(compact_level_ptr)}"
+            )
 
             # --- Write CSR HNSW graph data using unified function ---
-            print(f"[{time.time() - start_time:.2f}s] Writing CSR HNSW graph data in FAISS-compatible order...")
-            
+            print(
+                f"[{time.time() - start_time:.2f}s] Writing CSR HNSW graph data in FAISS-compatible order..."
+            )
+
             # Determine storage fourcc and data based on prune_embeddings
             if prune_embeddings:
-                print(f"   Pruning embeddings: Writing NULL storage marker.")
+                print("   Pruning embeddings: Writing NULL storage marker.")
                 output_storage_fourcc = NULL_INDEX_FOURCC
-                storage_data = b''
+                storage_data = b""
             else:
                 # Keep embeddings - read and preserve original storage data
                 if storage_fourcc and storage_fourcc != NULL_INDEX_FOURCC:
-                    print(f"   Preserving embeddings: Reading original storage data...")
+                    print("   Preserving embeddings: Reading original storage data...")
                     storage_data = f_in.read()  # Read remaining storage data
                     output_storage_fourcc = storage_fourcc
                     print(f"   Read {len(storage_data)} bytes of storage data")
                 else:
-                    print(f"   No embeddings found in original file (NULL storage)")
+                    print("   No embeddings found in original file (NULL storage)")
                     output_storage_fourcc = NULL_INDEX_FOURCC
-                    storage_data = b''
-            
+                    storage_data = b""
+
             # Use the unified write function
-            write_compact_format(f_out, original_hnsw_data, assign_probas_np, cum_nneighbor_per_level_np, 
-                               levels_np, compact_level_ptr, compact_node_offsets_np, 
-                               compact_neighbors_data, output_storage_fourcc, storage_data)
-            
+            write_compact_format(
+                f_out,
+                original_hnsw_data,
+                assign_probas_np,
+                cum_nneighbor_per_level_np,
+                levels_np,
+                compact_level_ptr,
+                compact_node_offsets_np,
+                compact_neighbors_data,
+                output_storage_fourcc,
+                storage_data,
+            )
+
             # Clean up memory
             del assign_probas_np, cum_nneighbor_per_level_np, levels_np
             del compact_neighbors_data, compact_level_ptr, compact_node_offsets_np
@@ -503,40 +651,66 @@ def convert_hnsw_graph_to_csr(input_filename, output_filename, prune_embeddings=
         print(f"Error: Input file not found: {input_filename}", file=sys.stderr)
         return False
     except MemoryError as e:
-         print(f"\nFatal MemoryError during conversion: {e}. Insufficient RAM.", file=sys.stderr)
-         # Clean up potentially partially written output file?
-         try: os.remove(output_filename)
-         except OSError: pass
-         return False
+        print(
+            f"\nFatal MemoryError during conversion: {e}. Insufficient RAM.",
+            file=sys.stderr,
+        )
+        # Clean up potentially partially written output file?
+        try:
+            os.remove(output_filename)
+        except OSError:
+            pass
+        return False
     except EOFError as e:
-        print(f"Error: Reached end of file unexpectedly reading {input_filename}. {e}", file=sys.stderr)
-        try: os.remove(output_filename)
-        except OSError: pass
+        print(
+            f"Error: Reached end of file unexpectedly reading {input_filename}. {e}",
+            file=sys.stderr,
+        )
+        try:
+            os.remove(output_filename)
+        except OSError:
+            pass
         return False
     except Exception as e:
         print(f"An unexpected error occurred during conversion: {e}", file=sys.stderr)
         import traceback
+
         traceback.print_exc()
         try:
             os.remove(output_filename)
-        except OSError: pass
+        except OSError:
+            pass
         return False
     # Ensure neighbors_np is deleted even if an error occurs after its allocation
     finally:
-        if 'neighbors_np' in locals() and neighbors_np is not None:
-            del neighbors_np
-            gc.collect()
+        try:
+            if "neighbors_np" in locals() and neighbors_np is not None:
+                del neighbors_np
+                gc.collect()
+        except NameError:
+            pass
 
 
 # --- Script Execution ---
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Convert a Faiss IndexHNSWFlat file to a CSR-based HNSW graph file.")
+    parser = argparse.ArgumentParser(
+        description="Convert a Faiss IndexHNSWFlat file to a CSR-based HNSW graph file."
+    )
     parser.add_argument("input_index_file", help="Path to the input IndexHNSWFlat file")
-    parser.add_argument("output_csr_graph_file", help="Path to write the output CSR HNSW graph file")
-    parser.add_argument("--prune-embeddings", action="store_true", default=True, 
-                       help="Prune embedding storage (write NULL storage marker)")
-    parser.add_argument("--keep-embeddings", action="store_true", 
-                       help="Keep embedding storage (overrides --prune-embeddings)")
+    parser.add_argument(
+        "output_csr_graph_file", help="Path to write the output CSR HNSW graph file"
+    )
+    parser.add_argument(
+        "--prune-embeddings",
+        action="store_true",
+        default=True,
+        help="Prune embedding storage (write NULL storage marker)",
+    )
+    parser.add_argument(
+        "--keep-embeddings",
+        action="store_true",
+        help="Keep embedding storage (overrides --prune-embeddings)",
+    )
 
     args = parser.parse_args()
 
@@ -545,10 +719,12 @@ if __name__ == "__main__":
         sys.exit(1)
 
     if os.path.abspath(args.input_index_file) == os.path.abspath(args.output_csr_graph_file):
-         print(f"Error: Input and output filenames cannot be the same.", file=sys.stderr)
-         sys.exit(1)
+        print("Error: Input and output filenames cannot be the same.", file=sys.stderr)
+        sys.exit(1)
 
     prune_embeddings = args.prune_embeddings and not args.keep_embeddings
-    success = convert_hnsw_graph_to_csr(args.input_index_file, args.output_csr_graph_file, prune_embeddings)
+    success = convert_hnsw_graph_to_csr(
+        args.input_index_file, args.output_csr_graph_file, prune_embeddings
+    )
     if not success:
-        sys.exit(1)
+        sys.exit(1)

packages/leann-backend-hnsw/leann_backend_hnsw/hnsw_backend.py

@@ -1,20 +1,21 @@
-import numpy as np
+import logging
 import os
-from pathlib import Path
-from typing import Dict, Any, List, Literal
-import pickle
 import shutil
-import time
+from pathlib import Path
+from typing import Any, Literal
 
-from leann.searcher_base import BaseSearcher
-from .convert_to_csr import convert_hnsw_graph_to_csr
-
-from leann.registry import register_backend
+import numpy as np
 from leann.interface import (
-    LeannBackendFactoryInterface,
     LeannBackendBuilderInterface,
+    LeannBackendFactoryInterface,
     LeannBackendSearcherInterface,
 )
+from leann.registry import register_backend
+from leann.searcher_base import BaseSearcher
+
+from .convert_to_csr import convert_hnsw_graph_to_csr
+
+logger = logging.getLogger(__name__)
 
 
 def get_metric_map():
@@ -27,6 +28,12 @@ def get_metric_map():
     }
 
 
+def normalize_l2(data: np.ndarray) -> np.ndarray:
+    norms = np.linalg.norm(data, axis=1, keepdims=True)
+    norms[norms == 0] = 1  # Avoid division by zero
+    return data / norms
+
+
 @register_backend("hnsw")
 class HNSWBackend(LeannBackendFactoryInterface):
     @staticmethod
@@ -47,8 +54,14 @@ class HNSWBuilder(LeannBackendBuilderInterface):
         self.efConstruction = self.build_params.setdefault("efConstruction", 200)
         self.distance_metric = self.build_params.setdefault("distance_metric", "mips")
         self.dimensions = self.build_params.get("dimensions")
+        if not self.is_recompute:
+            if self.is_compact:
+                # TODO: support this case @andy
+                raise ValueError(
+                    "is_recompute is False, but is_compact is True. This is not compatible now. change is compact to False and you can use the original HNSW index."
+                )
 
-    def build(self, data: np.ndarray, ids: List[str], index_path: str, **kwargs):
+    def build(self, data: np.ndarray, ids: list[str], index_path: str, **kwargs):
         from . import faiss  # type: ignore
 
         path = Path(index_path)
@@ -57,9 +70,9 @@ class HNSWBuilder(LeannBackendBuilderInterface):
         index_dir.mkdir(parents=True, exist_ok=True)
 
         if data.dtype != np.float32:
+            logger.warning(f"Converting data to float32, shape: {data.shape}")
             data = data.astype(np.float32)
 
-
         metric_enum = get_metric_map().get(self.distance_metric.lower())
         if metric_enum is None:
             raise ValueError(f"Unsupported distance_metric '{self.distance_metric}'.")
@@ -69,7 +82,7 @@ class HNSWBuilder(LeannBackendBuilderInterface):
         index.hnsw.efConstruction = self.efConstruction
 
         if self.distance_metric.lower() == "cosine":
-            faiss.normalize_L2(data)
+            data = normalize_l2(data)
 
         index.add(data.shape[0], faiss.swig_ptr(data))
         index_file = index_dir / f"{index_prefix}.index"
@@ -81,7 +94,7 @@ class HNSWBuilder(LeannBackendBuilderInterface):
     def _convert_to_csr(self, index_file: Path):
         """Convert built index to CSR format"""
         mode_str = "CSR-pruned" if self.is_recompute else "CSR-standard"
-        print(f"INFO: Converting HNSW index to {mode_str} format...")
+        logger.info(f"INFO: Converting HNSW index to {mode_str} format...")
 
         csr_temp_file = index_file.with_suffix(".csr.tmp")
 
@@ -90,20 +103,16 @@ class HNSWBuilder(LeannBackendBuilderInterface):
         )
 
         if success:
-            print("✅ CSR conversion successful.")
-            index_file_old = index_file.with_suffix(".old")
-            shutil.move(str(index_file), str(index_file_old))
+            logger.info("✅ CSR conversion successful.")
+            # index_file_old = index_file.with_suffix(".old")
+            # shutil.move(str(index_file), str(index_file_old))
             shutil.move(str(csr_temp_file), str(index_file))
-            print(
-                f"INFO: Replaced original index with {mode_str} version at '{index_file}'"
-            )
+            logger.info(f"INFO: Replaced original index with {mode_str} version at '{index_file}'")
         else:
             # Clean up and fail fast
             if csr_temp_file.exists():
                 os.remove(csr_temp_file)
-            raise RuntimeError(
-                "CSR conversion failed - cannot proceed with compact format"
-            )
+            raise RuntimeError("CSR conversion failed - cannot proceed with compact format")
 
 
 class HNSWSearcher(BaseSearcher):
@@ -115,7 +124,9 @@ class HNSWSearcher(BaseSearcher):
         )
         from . import faiss  # type: ignore
 
-        self.distance_metric = self.meta.get("distance_metric", "mips").lower()
+        self.distance_metric = (
+            self.meta.get("backend_kwargs", {}).get("distance_metric", "mips").lower()
+        )
         metric_enum = get_metric_map().get(self.distance_metric)
         if metric_enum is None:
             raise ValueError(f"Unsupported distance_metric '{self.distance_metric}'.")
@@ -131,27 +142,25 @@ class HNSWSearcher(BaseSearcher):
 
         hnsw_config = faiss.HNSWIndexConfig()
         hnsw_config.is_compact = self.is_compact
-        hnsw_config.is_recompute = self.is_pruned or kwargs.get("is_recompute", False)
-
-        if self.is_pruned and not hnsw_config.is_recompute:
-            raise RuntimeError("Index is pruned but recompute is disabled.")
+        hnsw_config.is_recompute = (
+            self.is_pruned
+        )  # In C++ code, it's called is_recompute, but it's only for loading IIUC.
 
         self._index = faiss.read_index(str(index_file), faiss.IO_FLAG_MMAP, hnsw_config)
-        
 
     def search(
         self,
         query: np.ndarray,
         top_k: int,
+        zmq_port: int | None = None,
         complexity: int = 64,
         beam_width: int = 1,
         prune_ratio: float = 0.0,
-        recompute_embeddings: bool = False,
+        recompute_embeddings: bool = True,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: int = 5557,
         batch_size: int = 0,
         **kwargs,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         """
         Search for nearest neighbors using HNSW index.
 
@@ -166,7 +175,7 @@ class HNSWSearcher(BaseSearcher):
                 - "global": Use global PQ queue size for selection (default)
                 - "local": Local pruning, sort and select best candidates
                 - "proportional": Base selection on new neighbor count ratio
-            zmq_port: ZMQ port for embedding server
+            zmq_port: ZMQ port for embedding server communication. Must be provided if recompute_embeddings is True.
             batch_size: Neighbor processing batch size, 0=disabled (HNSW-specific)
             **kwargs: Additional HNSW-specific parameters (for legacy compatibility)
 
@@ -175,26 +184,34 @@ class HNSWSearcher(BaseSearcher):
         """
         from . import faiss  # type: ignore
 
-        # Use recompute_embeddings parameter
-        use_recompute = recompute_embeddings or self.is_pruned
-        if use_recompute:
-            meta_file_path = self.index_dir / f"{self.index_path.name}.meta.json"
-            if not meta_file_path.exists():
-                raise RuntimeError(
-                    f"FATAL: Recompute enabled but metadata file not found: {meta_file_path}"
-                )
-            self._ensure_server_running(str(meta_file_path), port=zmq_port, **kwargs)
+        if not recompute_embeddings:
+            if self.is_pruned:
+                raise RuntimeError("Recompute is required for pruned index.")
+        if recompute_embeddings:
+            if zmq_port is None:
+                raise ValueError("zmq_port must be provided if recompute_embeddings is True")
 
         if query.dtype != np.float32:
             query = query.astype(np.float32)
         if self.distance_metric == "cosine":
-            faiss.normalize_L2(query)
+            query = normalize_l2(query)
 
         params = faiss.SearchParametersHNSW()
-        params.zmq_port = zmq_port
+        if zmq_port is not None:
+            params.zmq_port = zmq_port  # C++ code won't use this if recompute_embeddings is False
         params.efSearch = complexity
         params.beam_size = beam_width
 
+        # For OpenAI embeddings with cosine distance, disable relative distance check
+        # This prevents early termination when all scores are in a narrow range
+        embedding_model = self.meta.get("embedding_model", "").lower()
+        if self.distance_metric == "cosine" and any(
+            openai_model in embedding_model for openai_model in ["text-embedding", "openai"]
+        ):
+            params.check_relative_distance = False
+        else:
+            params.check_relative_distance = True
+
         # PQ pruning: direct mapping to HNSW's pq_pruning_ratio
         params.pq_pruning_ratio = prune_ratio
 
@@ -204,9 +221,7 @@ class HNSWSearcher(BaseSearcher):
             params.send_neigh_times_ratio = 0.0
         elif pruning_strategy == "proportional":
             params.local_prune = False
-            params.send_neigh_times_ratio = (
-                1.0  # Any value > 1e-6 triggers proportional mode
-            )
+            params.send_neigh_times_ratio = 1.0  # Any value > 1e-6 triggers proportional mode
         else:  # "global"
             params.local_prune = False
             params.send_neigh_times_ratio = 0.0
@@ -227,9 +242,6 @@ class HNSWSearcher(BaseSearcher):
             params,
         )
 
-        string_labels = [
-            [str(int_label) for int_label in batch_labels]
-            for batch_labels in labels
-        ]
+        string_labels = [[str(int_label) for int_label in batch_labels] for batch_labels in labels]
 
         return {"labels": string_labels, "distances": distances}

packages/leann-backend-hnsw/leann_backend_hnsw/hnsw_embedding_server.py

@@ -3,33 +3,37 @@ HNSW-specific embedding server
 """
 
 import argparse
+import json
+import logging
+import os
+import sys
 import threading
 import time
-import os
-import zmq
-import numpy as np
-import msgpack
-import json
 from pathlib import Path
-from typing import Dict, Any, Optional, Union
-import sys
-import logging
 
-RED = "\033[91m"
-RESET = "\033[0m"
+import msgpack
+import numpy as np
+import zmq
 
 # Set up logging based on environment variable
-LOG_LEVEL = os.getenv("LEANN_LOG_LEVEL", "INFO").upper()
-logging.basicConfig(
-    level=getattr(logging, LOG_LEVEL, logging.INFO),
-    format="%(asctime)s - %(levelname)s - %(message)s",
-)
+LOG_LEVEL = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
 logger = logging.getLogger(__name__)
 
+# Force set logger level (don't rely on basicConfig in subprocess)
+log_level = getattr(logging, LOG_LEVEL, logging.WARNING)
+logger.setLevel(log_level)
+
+# Ensure we have a handler if none exists
+if not logger.handlers:
+    handler = logging.StreamHandler()
+    formatter = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
+    handler.setFormatter(formatter)
+    logger.addHandler(handler)
+    logger.propagate = False
+
 
 def create_hnsw_embedding_server(
-    passages_file: Optional[str] = None,
-    passages_data: Optional[Dict[str, str]] = None,
+    passages_file: str | None = None,
     zmq_port: int = 5555,
     model_name: str = "sentence-transformers/all-mpnet-base-v2",
     distance_metric: str = "mips",
@@ -39,14 +43,8 @@ def create_hnsw_embedding_server(
     Create and start a ZMQ-based embedding server for HNSW backend.
     Simplified version using unified embedding computation module.
     """
-    # Auto-detect mode based on model name if not explicitly set
-    if embedding_mode == "sentence-transformers" and model_name.startswith(
-        "text-embedding-"
-    ):
-        embedding_mode = "openai"
-
-    print(f"Starting HNSW server on port {zmq_port} with model {model_name}")
-    print(f"Using embedding mode: {embedding_mode}")
+    logger.info(f"Starting HNSW server on port {zmq_port} with model {model_name}")
+    logger.info(f"Using embedding mode: {embedding_mode}")
 
     # Add leann-core to path for unified embedding computation
     current_dir = Path(__file__).parent
@@ -54,12 +52,12 @@ def create_hnsw_embedding_server(
     sys.path.insert(0, str(leann_core_path))
 
     try:
-        from leann.embedding_compute import compute_embeddings
         from leann.api import PassageManager
+        from leann.embedding_compute import compute_embeddings
 
-        print("Successfully imported unified embedding computation module")
+        logger.info("Successfully imported unified embedding computation module")
     except ImportError as e:
-        print(f"ERROR: Failed to import embedding computation module: {e}")
+        logger.error(f"Failed to import embedding computation module: {e}")
         return
     finally:
         sys.path.pop(0)
@@ -72,26 +70,40 @@ def create_hnsw_embedding_server(
             return s.connect_ex(("localhost", port)) == 0
 
     if check_port(zmq_port):
-        print(f"{RED}Port {zmq_port} is already in use{RESET}")
+        logger.error(f"Port {zmq_port} is already in use")
         return
 
     # Only support metadata file, fail fast for everything else
     if not passages_file or not passages_file.endswith(".meta.json"):
         raise ValueError("Only metadata files (.meta.json) are supported")
-    
+
     # Load metadata to get passage sources
-    with open(passages_file, "r") as f:
+    with open(passages_file) as f:
         meta = json.load(f)
-    
-    passages = PassageManager(meta["passage_sources"])
-    print(f"Loaded PassageManager with {len(passages.global_offset_map)} passages from metadata")
+
+    # Convert relative paths to absolute paths based on metadata file location
+    metadata_dir = Path(passages_file).parent.parent  # Go up one level from the metadata file
+    passage_sources = []
+    for source in meta["passage_sources"]:
+        source_copy = source.copy()
+        # Convert relative paths to absolute paths
+        if not Path(source_copy["path"]).is_absolute():
+            source_copy["path"] = str(metadata_dir / source_copy["path"])
+        if not Path(source_copy["index_path"]).is_absolute():
+            source_copy["index_path"] = str(metadata_dir / source_copy["index_path"])
+        passage_sources.append(source_copy)
+
+    passages = PassageManager(passage_sources)
+    logger.info(
+        f"Loaded PassageManager with {len(passages.global_offset_map)} passages from metadata"
+    )
 
     def zmq_server_thread():
         """ZMQ server thread"""
         context = zmq.Context()
         socket = context.socket(zmq.REP)
         socket.bind(f"tcp://*:{zmq_port}")
-        print(f"HNSW ZMQ server listening on port {zmq_port}")
+        logger.info(f"HNSW ZMQ server listening on port {zmq_port}")
 
         socket.setsockopt(zmq.RCVTIMEO, 300000)
         socket.setsockopt(zmq.SNDTIMEO, 300000)
@@ -99,12 +111,12 @@ def create_hnsw_embedding_server(
         while True:
             try:
                 message_bytes = socket.recv()
-                print(f"Received ZMQ request of size {len(message_bytes)} bytes")
+                logger.debug(f"Received ZMQ request of size {len(message_bytes)} bytes")
 
                 e2e_start = time.time()
                 request_payload = msgpack.unpackb(message_bytes)
 
-                # Handle direct text embedding request (for OpenAI and sentence-transformers)
+                # Handle direct text embedding request
                 if isinstance(request_payload, list) and len(request_payload) > 0:
                     # Check if this is a direct text request (list of strings)
                     if all(isinstance(item, str) for item in request_payload):
@@ -112,7 +124,7 @@ def create_hnsw_embedding_server(
                             f"Processing direct text embedding request for {len(request_payload)} texts in {embedding_mode} mode"
                         )
 
-                        # Use unified embedding computation
+                        # Use unified embedding computation (now with model caching)
                         embeddings = compute_embeddings(
                             request_payload, model_name, mode=embedding_mode
                         )
@@ -120,9 +132,7 @@ def create_hnsw_embedding_server(
                         response = embeddings.tolist()
                         socket.send(msgpack.packb(response))
                         e2e_end = time.time()
-                        logger.info(
-                            f"⏱️  Text embedding E2E time: {e2e_end - e2e_start:.6f}s"
-                        )
+                        logger.info(f"⏱️  Text embedding E2E time: {e2e_end - e2e_start:.6f}s")
                         continue
 
                 # Handle distance calculation requests
@@ -136,8 +146,8 @@ def create_hnsw_embedding_server(
                     query_vector = np.array(request_payload[1], dtype=np.float32)
 
                     logger.debug("Distance calculation request received")
-                    print(f"    Node IDs: {node_ids}")
-                    print(f"    Query vector dim: {len(query_vector)}")
+                    logger.debug(f"    Node IDs: {node_ids}")
+                    logger.debug(f"    Query vector dim: {len(query_vector)}")
 
                     # Get embeddings for node IDs
                     texts = []
@@ -147,18 +157,16 @@ def create_hnsw_embedding_server(
                             txt = passage_data["text"]
                             texts.append(txt)
                         except KeyError:
-                            print(f"ERROR: Passage ID {nid} not found")
+                            logger.error(f"Passage ID {nid} not found")
                             raise RuntimeError(f"FATAL: Passage with ID {nid} not found")
                         except Exception as e:
-                            print(f"ERROR: Exception looking up passage ID {nid}: {e}")
+                            logger.error(f"Exception looking up passage ID {nid}: {e}")
                             raise
 
                     # Process embeddings
-                    embeddings = compute_embeddings(
-                        texts, model_name, mode=embedding_mode
-                    )
-                    print(
-                        f"INFO: Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
+                    embeddings = compute_embeddings(texts, model_name, mode=embedding_mode)
+                    logger.info(
+                        f"Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
                     )
 
                     # Calculate distances
@@ -170,16 +178,12 @@ def create_hnsw_embedding_server(
                         distances = -np.dot(embeddings, query_vector)
 
                     response_payload = distances.flatten().tolist()
-                    response_bytes = msgpack.packb(
-                        [response_payload], use_single_float=True
-                    )
-                    print(f"Sending distance response with {len(distances)} distances")
+                    response_bytes = msgpack.packb([response_payload], use_single_float=True)
+                    logger.debug(f"Sending distance response with {len(distances)} distances")
 
                     socket.send(response_bytes)
                     e2e_end = time.time()
-                    logger.info(
-                        f"⏱️  Distance calculation E2E time: {e2e_end - e2e_start:.6f}s"
-                    )
+                    logger.info(f"⏱️  Distance calculation E2E time: {e2e_end - e2e_start:.6f}s")
                     continue
 
                 # Standard embedding request (passage ID lookup)
@@ -188,14 +192,14 @@ def create_hnsw_embedding_server(
                     or len(request_payload) != 1
                     or not isinstance(request_payload[0], list)
                 ):
-                    print(
-                        f"Error: Invalid MessagePack request format. Expected [[ids...]] or [texts...], got: {type(request_payload)}"
+                    logger.error(
+                        f"Invalid MessagePack request format. Expected [[ids...]] or [texts...], got: {type(request_payload)}"
                     )
                     socket.send(msgpack.packb([[], []]))
                     continue
 
                 node_ids = request_payload[0]
-                print(f"Request for {len(node_ids)} node embeddings")
+                logger.debug(f"Request for {len(node_ids)} node embeddings")
 
                 # Look up texts by node IDs
                 texts = []
@@ -209,25 +213,23 @@ def create_hnsw_embedding_server(
                     except KeyError:
                         raise RuntimeError(f"FATAL: Passage with ID {nid} not found")
                     except Exception as e:
-                        print(f"ERROR: Exception looking up passage ID {nid}: {e}")
+                        logger.error(f"Exception looking up passage ID {nid}: {e}")
                         raise
 
                 # Process embeddings
                 embeddings = compute_embeddings(texts, model_name, mode=embedding_mode)
-                print(
-                    f"INFO: Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
+                logger.info(
+                    f"Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
                 )
 
                 # Serialization and response
                 if np.isnan(embeddings).any() or np.isinf(embeddings).any():
-                    print(
-                        f"{RED}!!! ERROR: NaN or Inf detected in embeddings! Requested IDs: {node_ids[:5]}...{RESET}"
+                    logger.error(
+                        f"NaN or Inf detected in embeddings! Requested IDs: {node_ids[:5]}..."
                     )
-                    assert False
+                    raise AssertionError()
 
-                hidden_contiguous_f32 = np.ascontiguousarray(
-                    embeddings, dtype=np.float32
-                )
+                hidden_contiguous_f32 = np.ascontiguousarray(embeddings, dtype=np.float32)
                 response_payload = [
                     list(hidden_contiguous_f32.shape),
                     hidden_contiguous_f32.flatten().tolist(),
@@ -242,7 +244,7 @@ def create_hnsw_embedding_server(
                 logger.debug("ZMQ socket timeout, continuing to listen")
                 continue
             except Exception as e:
-                print(f"Error in ZMQ server loop: {e}")
+                logger.error(f"Error in ZMQ server loop: {e}")
                 import traceback
 
                 traceback.print_exc()
@@ -250,18 +252,29 @@ def create_hnsw_embedding_server(
 
     zmq_thread = threading.Thread(target=zmq_server_thread, daemon=True)
     zmq_thread.start()
-    print(f"Started HNSW ZMQ server thread on port {zmq_port}")
+    logger.info(f"Started HNSW ZMQ server thread on port {zmq_port}")
 
     # Keep the main thread alive
     try:
         while True:
             time.sleep(1)
     except KeyboardInterrupt:
-        print("HNSW Server shutting down...")
+        logger.info("HNSW Server shutting down...")
         return
 
 
 if __name__ == "__main__":
+    import signal
+    import sys
+
+    def signal_handler(sig, frame):
+        logger.info(f"Received signal {sig}, shutting down gracefully...")
+        sys.exit(0)
+
+    # Register signal handlers for graceful shutdown
+    signal.signal(signal.SIGTERM, signal_handler)
+    signal.signal(signal.SIGINT, signal_handler)
+
     parser = argparse.ArgumentParser(description="HNSW Embedding service")
     parser.add_argument("--zmq-port", type=int, default=5555, help="ZMQ port to run on")
     parser.add_argument(
@@ -282,7 +295,7 @@ if __name__ == "__main__":
         "--embedding-mode",
         type=str,
         default="sentence-transformers",
-        choices=["sentence-transformers", "openai"],
+        choices=["sentence-transformers", "openai", "mlx"],
         help="Embedding backend mode",
     )
 

packages/leann-backend-hnsw/pyproject.toml

@@ -6,9 +6,14 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "leann-backend-hnsw"
-version = "0.1.0"
+version = "0.1.16"
 description = "Custom-built HNSW (Faiss) backend for the Leann toolkit."
-dependencies = ["leann-core==0.1.0", "numpy"]
+dependencies = [
+    "leann-core==0.1.16",
+    "numpy",
+    "pyzmq>=23.0.0",
+    "msgpack>=1.0.0",
+]
 
 [tool.scikit-build]
 wheel.packages = ["leann_backend_hnsw"]
@@ -19,4 +24,4 @@ build.tool-args = ["-j8"]
 
 # CMake definitions to optimize compilation
 [tool.scikit-build.cmake.define]
-CMAKE_BUILD_PARALLEL_LEVEL = "8"
+CMAKE_BUILD_PARALLEL_LEVEL = "8"

packages/leann-core/pyproject.toml

@@ -4,19 +4,46 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "leann-core"
-version = "0.1.0"
-description = "Core API and plugin system for Leann."
+version = "0.1.16"
+description = "Core API and plugin system for LEANN"
 readme = "README.md"
 requires-python = ">=3.9"
 license = { text = "MIT" }
 
+# All required dependencies included
 dependencies = [
     "numpy>=1.20.0",
-    "tqdm>=4.60.0"
+    "tqdm>=4.60.0",
+    "psutil>=5.8.0",
+    "pyzmq>=23.0.0",
+    "msgpack>=1.0.0",
+    "torch>=2.0.0",
+    "sentence-transformers>=2.2.0",
+    "llama-index-core>=0.12.0",
+    "llama-index-readers-file>=0.4.0",  # Essential for document reading
+    "llama-index-embeddings-huggingface>=0.5.5",  # For embeddings
+    "python-dotenv>=1.0.0",
+    "openai>=1.0.0",
+    "huggingface-hub>=0.20.0",
+    "transformers>=4.30.0",
+    "requests>=2.25.0",
+    "accelerate>=0.20.0",
+    "PyPDF2>=3.0.0",
+    "pymupdf>=1.23.0",
+    "pdfplumber>=0.10.0",
+    "mlx>=0.26.3; sys_platform == 'darwin'",
+    "mlx-lm>=0.26.0; sys_platform == 'darwin'",
+]
+
+[project.optional-dependencies]
+colab = [
+    "torch>=2.0.0,<3.0.0",  # Limit torch version to avoid conflicts
+    "transformers>=4.30.0,<5.0.0",  # Limit transformers version
+    "accelerate>=0.20.0,<1.0.0",  # Limit accelerate version
 ]
 
 [project.scripts]
 leann = "leann.cli:main"
 
 [tool.setuptools.packages.find]
-where = ["src"]
+where = ["src"]

packages/leann-core/src/leann/__init__.py

@@ -8,10 +8,14 @@ if platform.system() == "Darwin":
     os.environ["MKL_NUM_THREADS"] = "1"
     os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
     os.environ["KMP_BLOCKTIME"] = "0"
+    # Additional fixes for PyTorch/sentence-transformers on macOS ARM64 only in CI
+    if os.environ.get("CI") == "true":
+        os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "0"
+        os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
 from .api import LeannBuilder, LeannChat, LeannSearcher
 from .registry import BACKEND_REGISTRY, autodiscover_backends
 
 autodiscover_backends()
 
-__all__ = ["LeannBuilder", "LeannSearcher", "LeannChat", "BACKEND_REGISTRY"]
+__all__ = ["BACKEND_REGISTRY", "LeannBuilder", "LeannChat", "LeannSearcher"]

packages/leann-core/src/leann/api.py

@@ -4,22 +4,37 @@ with the correct, original embedding logic from the user's reference code.
 """
 
 import json
+import logging
 import pickle
-import numpy as np
-from pathlib import Path
-from typing import List, Dict, Any, Optional, Literal
+import time
+import warnings
 from dataclasses import dataclass, field
-from .registry import BACKEND_REGISTRY
-from .interface import LeannBackendFactoryInterface
+from pathlib import Path
+from typing import Any, Literal
+
+import numpy as np
+
+from leann.interface import LeannBackendSearcherInterface
+
 from .chat import get_llm
+from .interface import LeannBackendFactoryInterface
+from .registry import BACKEND_REGISTRY
+
+logger = logging.getLogger(__name__)
+
+
+def get_registered_backends() -> list[str]:
+    """Get list of registered backend names."""
+    return list(BACKEND_REGISTRY.keys())
 
 
 def compute_embeddings(
-    chunks: List[str],
+    chunks: list[str],
     model_name: str,
     mode: str = "sentence-transformers",
     use_server: bool = True,
-    port: int = 5557,
+    port: int | None = None,
+    is_build=False,
 ) -> np.ndarray:
     """
     Computes embeddings using different backends.
@@ -38,6 +53,8 @@ def compute_embeddings(
     """
     if use_server:
         # Use embedding server (for search/query)
+        if port is None:
+            raise ValueError("port is required when use_server is True")
         return compute_embeddings_via_server(chunks, model_name, port=port)
     else:
         # Use direct computation (for build_index)
@@ -49,24 +66,23 @@ def compute_embeddings(
             chunks,
             model_name,
             mode=mode,
+            is_build=is_build,
         )
 
 
-def compute_embeddings_via_server(
-    chunks: List[str], model_name: str, port: int
-) -> np.ndarray:
+def compute_embeddings_via_server(chunks: list[str], model_name: str, port: int) -> np.ndarray:
     """Computes embeddings using sentence-transformers.
 
     Args:
         chunks: List of text chunks to embed
         model_name: Name of the sentence transformer model
     """
-    print(
-        f"INFO: Computing embeddings for {len(chunks)} chunks using SentenceTransformer model '{model_name}' (via embedding server)..."
+    logger.info(
+        f"Computing embeddings for {len(chunks)} chunks using SentenceTransformer model '{model_name}' (via embedding server)..."
     )
-    import zmq
     import msgpack
     import numpy as np
+    import zmq
 
     # Connect to embedding server
     context = zmq.Context()
@@ -95,36 +111,42 @@ class SearchResult:
     id: str
     score: float
     text: str
-    metadata: Dict[str, Any] = field(default_factory=dict)
+    metadata: dict[str, Any] = field(default_factory=dict)
 
 
 class PassageManager:
-    def __init__(self, passage_sources: List[Dict[str, Any]]):
+    def __init__(self, passage_sources: list[dict[str, Any]]):
         self.offset_maps = {}
         self.passage_files = {}
         self.global_offset_map = {}  # Combined map for fast lookup
 
         for source in passage_sources:
-            if source["type"] == "jsonl":
-                passage_file = source["path"]
-                index_file = source["index_path"]
-                if not Path(index_file).exists():
-                    raise FileNotFoundError(
-                        f"Passage index file not found: {index_file}"
-                    )
-                with open(index_file, "rb") as f:
-                    offset_map = pickle.load(f)
-                    self.offset_maps[passage_file] = offset_map
-                    self.passage_files[passage_file] = passage_file
+            assert source["type"] == "jsonl", "only jsonl is supported"
+            passage_file = source["path"]
+            index_file = source["index_path"]  # .idx file
 
-                    # Build global map for O(1) lookup
-                    for passage_id, offset in offset_map.items():
-                        self.global_offset_map[passage_id] = (passage_file, offset)
+            # Fix path resolution for Colab and other environments
+            if not Path(index_file).is_absolute():
+                # If relative path, try to resolve it properly
+                index_file = str(Path(index_file).resolve())
 
-    def get_passage(self, passage_id: str) -> Dict[str, Any]:
+            if not Path(index_file).exists():
+                raise FileNotFoundError(f"Passage index file not found: {index_file}")
+
+            with open(index_file, "rb") as f:
+                offset_map = pickle.load(f)
+                self.offset_maps[passage_file] = offset_map
+                self.passage_files[passage_file] = passage_file
+
+                # Build global map for O(1) lookup
+                for passage_id, offset in offset_map.items():
+                    self.global_offset_map[passage_id] = (passage_file, offset)
+
+    def get_passage(self, passage_id: str) -> dict[str, Any]:
         if passage_id in self.global_offset_map:
             passage_file, offset = self.global_offset_map[passage_id]
-            with open(passage_file, "r", encoding="utf-8") as f:
+            # Lazy file opening - only open when needed
+            with open(passage_file, encoding="utf-8") as f:
                 f.seek(offset)
                 return json.loads(f.readline())
         raise KeyError(f"Passage ID not found: {passage_id}")
@@ -134,25 +156,93 @@ class LeannBuilder:
     def __init__(
         self,
         backend_name: str,
-        embedding_model: str = "facebook/contriever-msmarco",
-        dimensions: Optional[int] = None,
+        embedding_model: str = "facebook/contriever",
+        dimensions: int | None = None,
         embedding_mode: str = "sentence-transformers",
         **backend_kwargs,
     ):
         self.backend_name = backend_name
-        backend_factory: LeannBackendFactoryInterface | None = BACKEND_REGISTRY.get(
-            backend_name
-        )
+        backend_factory: LeannBackendFactoryInterface | None = BACKEND_REGISTRY.get(backend_name)
         if backend_factory is None:
             raise ValueError(f"Backend '{backend_name}' not found or not registered.")
         self.backend_factory = backend_factory
         self.embedding_model = embedding_model
         self.dimensions = dimensions
         self.embedding_mode = embedding_mode
-        self.backend_kwargs = backend_kwargs
-        self.chunks: List[Dict[str, Any]] = []
 
-    def add_text(self, text: str, metadata: Optional[Dict[str, Any]] = None):
+        # Check if we need to use cosine distance for normalized embeddings
+        normalized_embeddings_models = {
+            # OpenAI models
+            ("openai", "text-embedding-ada-002"),
+            ("openai", "text-embedding-3-small"),
+            ("openai", "text-embedding-3-large"),
+            # Voyage AI models
+            ("voyage", "voyage-2"),
+            ("voyage", "voyage-3"),
+            ("voyage", "voyage-large-2"),
+            ("voyage", "voyage-multilingual-2"),
+            ("voyage", "voyage-code-2"),
+            # Cohere models
+            ("cohere", "embed-english-v3.0"),
+            ("cohere", "embed-multilingual-v3.0"),
+            ("cohere", "embed-english-light-v3.0"),
+            ("cohere", "embed-multilingual-light-v3.0"),
+        }
+
+        # Also check for patterns in model names
+        is_normalized = False
+        current_model_lower = embedding_model.lower()
+        current_mode_lower = embedding_mode.lower()
+
+        # Check exact matches
+        for mode, model in normalized_embeddings_models:
+            if (current_mode_lower == mode and current_model_lower == model) or (
+                mode in current_mode_lower and model in current_model_lower
+            ):
+                is_normalized = True
+                break
+
+        # Check patterns
+        if not is_normalized:
+            # OpenAI patterns
+            if "openai" in current_mode_lower or "openai" in current_model_lower:
+                if any(
+                    pattern in current_model_lower
+                    for pattern in ["text-embedding", "ada", "3-small", "3-large"]
+                ):
+                    is_normalized = True
+            # Voyage patterns
+            elif "voyage" in current_mode_lower or "voyage" in current_model_lower:
+                is_normalized = True
+            # Cohere patterns
+            elif "cohere" in current_mode_lower or "cohere" in current_model_lower:
+                if "embed" in current_model_lower:
+                    is_normalized = True
+
+        # Handle distance metric
+        if is_normalized and "distance_metric" not in backend_kwargs:
+            backend_kwargs["distance_metric"] = "cosine"
+            warnings.warn(
+                f"Detected normalized embeddings model '{embedding_model}' with mode '{embedding_mode}'. "
+                f"Automatically setting distance_metric='cosine' for optimal performance. "
+                f"Normalized embeddings (L2 norm = 1) should use cosine similarity instead of MIPS.",
+                UserWarning,
+                stacklevel=2,
+            )
+        elif is_normalized and backend_kwargs.get("distance_metric", "").lower() != "cosine":
+            current_metric = backend_kwargs.get("distance_metric", "mips")
+            warnings.warn(
+                f"Warning: Using '{current_metric}' distance metric with normalized embeddings model "
+                f"'{embedding_model}' may lead to suboptimal search results. "
+                f"Consider using 'cosine' distance metric for better performance.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+        self.backend_kwargs = backend_kwargs
+        self.chunks: list[dict[str, Any]] = []
+
+    def add_text(self, text: str, metadata: dict[str, Any] | None = None):
         if metadata is None:
             metadata = {}
         passage_id = metadata.get("id", str(len(self.chunks)))
@@ -182,9 +272,7 @@ class LeannBuilder:
             try:
                 from tqdm import tqdm
 
-                chunk_iterator = tqdm(
-                    self.chunks, desc="Writing passages", unit="chunk"
-                )
+                chunk_iterator = tqdm(self.chunks, desc="Writing passages", unit="chunk")
             except ImportError:
                 chunk_iterator = self.chunks
 
@@ -209,14 +297,12 @@ class LeannBuilder:
             self.embedding_model,
             self.embedding_mode,
             use_server=False,
-            port=5557,
+            is_build=True,
         )
         string_ids = [chunk["id"] for chunk in self.chunks]
         current_backend_kwargs = {**self.backend_kwargs, "dimensions": self.dimensions}
         builder_instance = self.backend_factory.builder(**current_backend_kwargs)
-        builder_instance.build(
-            embeddings, string_ids, index_path, **current_backend_kwargs
-        )
+        builder_instance.build(embeddings, string_ids, index_path, **current_backend_kwargs)
         leann_meta_path = index_dir / f"{index_name}.meta.json"
         meta_data = {
             "version": "1.0",
@@ -265,9 +351,7 @@ class LeannBuilder:
         ids, embeddings = data
 
         if not isinstance(embeddings, np.ndarray):
-            raise ValueError(
-                f"Expected embeddings to be numpy array, got {type(embeddings)}"
-            )
+            raise ValueError(f"Expected embeddings to be numpy array, got {type(embeddings)}")
 
         if len(ids) != embeddings.shape[0]:
             raise ValueError(
@@ -279,11 +363,9 @@ class LeannBuilder:
         if self.dimensions is None:
             self.dimensions = embedding_dim
         elif self.dimensions != embedding_dim:
-            raise ValueError(
-                f"Dimension mismatch: expected {self.dimensions}, got {embedding_dim}"
-            )
+            raise ValueError(f"Dimension mismatch: expected {self.dimensions}, got {embedding_dim}")
 
-        print(
+        logger.info(
             f"Building index from precomputed embeddings: {len(ids)} items, {embedding_dim} dimensions"
         )
 
@@ -291,7 +373,7 @@ class LeannBuilder:
         if len(self.chunks) != len(ids):
             # If no text chunks provided, create placeholder text entries
             if not self.chunks:
-                print("No text chunks provided, creating placeholder entries...")
+                logger.info("No text chunks provided, creating placeholder entries...")
                 for id_val in ids:
                     self.add_text(
                         f"Document {id_val}",
@@ -366,32 +448,33 @@ class LeannBuilder:
         with open(leann_meta_path, "w", encoding="utf-8") as f:
             json.dump(meta_data, f, indent=2)
 
-        print(f"Index built successfully from precomputed embeddings: {index_path}")
+        logger.info(f"Index built successfully from precomputed embeddings: {index_path}")
 
 
 class LeannSearcher:
     def __init__(self, index_path: str, enable_warmup: bool = False, **backend_kwargs):
-        meta_path_str = f"{index_path}.meta.json"
-        if not Path(meta_path_str).exists():
-            raise FileNotFoundError(f"Leann metadata file not found at {meta_path_str}")
-        with open(meta_path_str, "r", encoding="utf-8") as f:
+        # Fix path resolution for Colab and other environments
+        if not Path(index_path).is_absolute():
+            index_path = str(Path(index_path).resolve())
+
+        self.meta_path_str = f"{index_path}.meta.json"
+        if not Path(self.meta_path_str).exists():
+            raise FileNotFoundError(f"Leann metadata file not found at {self.meta_path_str}")
+        with open(self.meta_path_str, encoding="utf-8") as f:
             self.meta_data = json.load(f)
         backend_name = self.meta_data["backend_name"]
         self.embedding_model = self.meta_data["embedding_model"]
         # Support both old and new format
-        self.embedding_mode = self.meta_data.get(
-            "embedding_mode", "sentence-transformers"
-        )
-        # Backward compatibility with use_mlx
-        if self.meta_data.get("use_mlx", False):
-            self.embedding_mode = "mlx"
+        self.embedding_mode = self.meta_data.get("embedding_mode", "sentence-transformers")
         self.passage_manager = PassageManager(self.meta_data.get("passage_sources", []))
         backend_factory = BACKEND_REGISTRY.get(backend_name)
         if backend_factory is None:
             raise ValueError(f"Backend '{backend_name}' not found.")
         final_kwargs = {**self.meta_data.get("backend_kwargs", {}), **backend_kwargs}
         final_kwargs["enable_warmup"] = enable_warmup
-        self.backend_impl = backend_factory.searcher(index_path, **final_kwargs)
+        self.backend_impl: LeannBackendSearcherInterface = backend_factory.searcher(
+            index_path, **final_kwargs
+        )
 
     def search(
         self,
@@ -400,26 +483,39 @@ class LeannSearcher:
         complexity: int = 64,
         beam_width: int = 1,
         prune_ratio: float = 0.0,
-        recompute_embeddings: bool = False,
+        recompute_embeddings: bool = True,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: int = 5557,
+        expected_zmq_port: int = 5557,
         **kwargs,
-    ) -> List[SearchResult]:
-        print("🔍 DEBUG LeannSearcher.search() called:")
-        print(f"  Query: '{query}'")
-        print(f"  Top_k: {top_k}")
-        print(f"  Additional kwargs: {kwargs}")
+    ) -> list[SearchResult]:
+        logger.info("🔍 LeannSearcher.search() called:")
+        logger.info(f"  Query: '{query}'")
+        logger.info(f"  Top_k: {top_k}")
+        logger.info(f"  Additional kwargs: {kwargs}")
 
-        # Use backend's compute_query_embedding method
-        # This will automatically use embedding server if available and needed
-        import time
+        zmq_port = None
+
+        start_time = time.time()
+        if recompute_embeddings:
+            zmq_port = self.backend_impl._ensure_server_running(
+                self.meta_path_str,
+                port=expected_zmq_port,
+                **kwargs,
+            )
+            del expected_zmq_port
+        zmq_time = time.time() - start_time
+        logger.info(f"  Launching server time: {zmq_time} seconds")
 
         start_time = time.time()
 
-        query_embedding = self.backend_impl.compute_query_embedding(query, zmq_port)
-        print(f"  Generated embedding shape: {query_embedding.shape}")
-        embedding_time = time.time() - start_time
-        print(f"  Embedding time: {embedding_time} seconds")
+        query_embedding = self.backend_impl.compute_query_embedding(
+            query,
+            use_server_if_available=recompute_embeddings,
+            zmq_port=zmq_port,
+        )
+        # logger.info(f"  Generated embedding shape: {query_embedding.shape}")
+        time.time() - start_time
+        # logger.info(f"  Embedding time: {embedding_time} seconds")
 
         start_time = time.time()
         results = self.backend_impl.search(
@@ -433,17 +529,15 @@ class LeannSearcher:
             zmq_port=zmq_port,
             **kwargs,
         )
-        search_time = time.time() - start_time
-        print(f"  Search time: {search_time} seconds")
-        print(
-            f"  Backend returned: labels={len(results.get('labels', [[]])[0])} results"
-        )
+        time.time() - start_time
+        # logger.info(f"  Search time: {search_time} seconds")
+        logger.info(f"  Backend returned: labels={len(results.get('labels', [[]])[0])} results")
 
         enriched_results = []
         if "labels" in results and "distances" in results:
-            print(f"  Processing {len(results['labels'][0])} passage IDs:")
+            logger.info(f"  Processing {len(results['labels'][0])} passage IDs:")
             for i, (string_id, dist) in enumerate(
-                zip(results["labels"][0], results["distances"][0])
+                zip(results["labels"][0], results["distances"][0], strict=False)
             ):
                 try:
                     passage_data = self.passage_manager.get_passage(string_id)
@@ -455,15 +549,25 @@ class LeannSearcher:
                             metadata=passage_data.get("metadata", {}),
                         )
                     )
-                    print(
-                        f"    {i + 1}. passage_id='{string_id}' -> SUCCESS: {passage_data['text']}..."
+
+                    # Color codes for better logging
+                    GREEN = "\033[92m"
+                    BLUE = "\033[94m"
+                    YELLOW = "\033[93m"
+                    RESET = "\033[0m"
+
+                    # Truncate text for display (first 100 chars)
+                    display_text = passage_data["text"]
+                    logger.info(
+                        f"   {GREEN}✓{RESET} {BLUE}[{i + 1:2d}]{RESET} {YELLOW}ID:{RESET} '{string_id}' {YELLOW}Score:{RESET} {dist:.4f} {YELLOW}Text:{RESET} {display_text}"
                     )
                 except KeyError:
-                    print(
-                        f"    {i + 1}. passage_id='{string_id}' -> ERROR: Passage not found in PassageManager!"
+                    RED = "\033[91m"
+                    logger.error(
+                        f"   {RED}✗{RESET} [{i + 1:2d}] ID: '{string_id}' -> {RED}ERROR: Passage not found!{RESET}"
                     )
 
-        print(f"  Final enriched results: {len(enriched_results)} passages")
+        logger.info(f"  {GREEN}✓ Final enriched results: {len(enriched_results)} passages{RESET}")
         return enriched_results
 
 
@@ -471,7 +575,7 @@ class LeannChat:
     def __init__(
         self,
         index_path: str,
-        llm_config: Optional[Dict[str, Any]] = None,
+        llm_config: dict[str, Any] | None = None,
         enable_warmup: bool = False,
         **kwargs,
     ):
@@ -485,15 +589,15 @@ class LeannChat:
         complexity: int = 64,
         beam_width: int = 1,
         prune_ratio: float = 0.0,
-        recompute_embeddings: bool = False,
+        recompute_embeddings: bool = True,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: int = 5557,
-        llm_kwargs: Optional[Dict[str, Any]] = None,
+        llm_kwargs: dict[str, Any] | None = None,
+        expected_zmq_port: int = 5557,
         **search_kwargs,
     ):
         if llm_kwargs is None:
             llm_kwargs = {}
-
+        search_time = time.time()
         results = self.searcher.search(
             question,
             top_k=top_k,
@@ -502,9 +606,11 @@ class LeannChat:
             prune_ratio=prune_ratio,
             recompute_embeddings=recompute_embeddings,
             pruning_strategy=pruning_strategy,
-            zmq_port=zmq_port,
+            expected_zmq_port=expected_zmq_port,
             **search_kwargs,
         )
+        search_time = time.time() - search_time
+        # logger.info(f"  Search time: {search_time} seconds")
         context = "\n\n".join([r.text for r in results])
         prompt = (
             "Here is some retrieved context that might help answer your question:\n\n"

packages/leann-core/src/leann/chat.py

@@ -4,21 +4,24 @@ This file contains the chat generation logic for the LEANN project,
 supporting different backends like Ollama, Hugging Face Transformers, and a simulation mode.
 """
 
-from abc import ABC, abstractmethod
-from typing import Dict, Any, Optional, List
+import difflib
 import logging
 import os
-import difflib
+from abc import ABC, abstractmethod
+from typing import Any
+
+import torch
 
 # Configure logging
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
 
-def check_ollama_models() -> List[str]:
+def check_ollama_models() -> list[str]:
     """Check available Ollama models and return a list"""
     try:
         import requests
+
         response = requests.get("http://localhost:11434/api/tags", timeout=5)
         if response.status_code == 200:
             data = response.json()
@@ -28,68 +31,135 @@ def check_ollama_models() -> List[str]:
         return []
 
 
-def search_ollama_models_fuzzy(query: str, available_models: List[str]) -> List[str]:
+def check_ollama_model_exists_remotely(model_name: str) -> tuple[bool, list[str]]:
+    """Check if a model exists in Ollama's remote library and return available tags
+
+    Returns:
+        (model_exists, available_tags): bool and list of matching tags
+    """
+    try:
+        import re
+
+        import requests
+
+        # Split model name and tag
+        if ":" in model_name:
+            base_model, requested_tag = model_name.split(":", 1)
+        else:
+            base_model, requested_tag = model_name, None
+
+        # First check if base model exists in library
+        library_response = requests.get("https://ollama.com/library", timeout=8)
+        if library_response.status_code != 200:
+            return True, []  # Assume exists if can't check
+
+        # Extract model names from library page
+        models_in_library = re.findall(r'href="/library/([^"]+)"', library_response.text)
+
+        if base_model not in models_in_library:
+            return False, []  # Base model doesn't exist
+
+        # If base model exists, get available tags
+        tags_response = requests.get(f"https://ollama.com/library/{base_model}/tags", timeout=8)
+        if tags_response.status_code != 200:
+            return True, []  # Base model exists but can't get tags
+
+        # Extract tags for this model - be more specific to avoid HTML artifacts
+        tag_pattern = rf"{re.escape(base_model)}:[a-zA-Z0-9\.\-_]+"
+        raw_tags = re.findall(tag_pattern, tags_response.text)
+
+        # Clean up tags - remove HTML artifacts and duplicates
+        available_tags = []
+        seen = set()
+        for tag in raw_tags:
+            # Skip if it looks like HTML (contains < or >)
+            if "<" in tag or ">" in tag:
+                continue
+            if tag not in seen:
+                seen.add(tag)
+                available_tags.append(tag)
+
+        # Check if exact model exists
+        if requested_tag is None:
+            # User just requested base model, suggest tags
+            return True, available_tags[:10]  # Return up to 10 tags
+        else:
+            exact_match = model_name in available_tags
+            return exact_match, available_tags[:10]
+
+    except Exception:
+        pass
+
+    # If scraping fails, assume model might exist (don't block user)
+    return True, []
+
+
+def search_ollama_models_fuzzy(query: str, available_models: list[str]) -> list[str]:
     """Use intelligent fuzzy search for Ollama models"""
     if not available_models:
         return []
-    
+
     query_lower = query.lower()
     suggestions = []
-    
+
     # 1. Exact matches first
     exact_matches = [m for m in available_models if query_lower == m.lower()]
     suggestions.extend(exact_matches)
-    
+
     # 2. Starts with query
-    starts_with = [m for m in available_models if m.lower().startswith(query_lower) and m not in suggestions]
+    starts_with = [
+        m for m in available_models if m.lower().startswith(query_lower) and m not in suggestions
+    ]
     suggestions.extend(starts_with)
-    
+
     # 3. Contains query
     contains = [m for m in available_models if query_lower in m.lower() and m not in suggestions]
     suggestions.extend(contains)
-    
+
     # 4. Base model name matching (remove version numbers)
     def get_base_name(model_name: str) -> str:
         """Extract base name without version (e.g., 'llama3:8b' -> 'llama3')"""
-        return model_name.split(':')[0].split('-')[0]
-    
+        return model_name.split(":")[0].split("-")[0]
+
     query_base = get_base_name(query_lower)
     base_matches = [
-        m for m in available_models 
+        m
+        for m in available_models
         if get_base_name(m.lower()) == query_base and m not in suggestions
     ]
     suggestions.extend(base_matches)
-    
+
     # 5. Family/variant matching
     model_families = {
-        'llama': ['llama2', 'llama3', 'alpaca', 'vicuna', 'codellama'],
-        'qwen': ['qwen', 'qwen2', 'qwen3'],
-        'gemma': ['gemma', 'gemma2'],
-        'phi': ['phi', 'phi2', 'phi3'],
-        'mistral': ['mistral', 'mixtral', 'openhermes'],
-        'dolphin': ['dolphin', 'openchat'],
-        'deepseek': ['deepseek', 'deepseek-coder']
+        "llama": ["llama2", "llama3", "alpaca", "vicuna", "codellama"],
+        "qwen": ["qwen", "qwen2", "qwen3"],
+        "gemma": ["gemma", "gemma2"],
+        "phi": ["phi", "phi2", "phi3"],
+        "mistral": ["mistral", "mixtral", "openhermes"],
+        "dolphin": ["dolphin", "openchat"],
+        "deepseek": ["deepseek", "deepseek-coder"],
     }
-    
+
     query_family = None
     for family, variants in model_families.items():
         if any(variant in query_lower for variant in variants):
             query_family = family
             break
-    
+
     if query_family:
         family_variants = model_families[query_family]
         family_matches = [
-            m for m in available_models
+            m
+            for m in available_models
             if any(variant in m.lower() for variant in family_variants) and m not in suggestions
         ]
         suggestions.extend(family_matches)
-    
+
     # 6. Use difflib for remaining fuzzy matches
     remaining_models = [m for m in available_models if m not in suggestions]
     difflib_matches = difflib.get_close_matches(query_lower, remaining_models, n=3, cutoff=0.4)
     suggestions.extend(difflib_matches)
-    
+
     return suggestions[:8]  # Return top 8 suggestions
 
 
@@ -99,15 +169,13 @@ def search_ollama_models_fuzzy(query: str, available_models: List[str]) -> List[
 # Remove this too - no need for fallback
 
 
-def suggest_similar_models(invalid_model: str, available_models: List[str]) -> List[str]:
+def suggest_similar_models(invalid_model: str, available_models: list[str]) -> list[str]:
     """Use difflib to find similar model names"""
     if not available_models:
         return []
-    
+
     # Get close matches using fuzzy matching
-    suggestions = difflib.get_close_matches(
-        invalid_model, available_models, n=3, cutoff=0.3
-    )
+    suggestions = difflib.get_close_matches(invalid_model, available_models, n=3, cutoff=0.3)
     return suggestions
 
 
@@ -115,49 +183,50 @@ def check_hf_model_exists(model_name: str) -> bool:
     """Quick check if HuggingFace model exists without downloading"""
     try:
         from huggingface_hub import model_info
+
         model_info(model_name)
         return True
     except Exception:
         return False
 
 
-def get_popular_hf_models() -> List[str]:
+def get_popular_hf_models() -> list[str]:
     """Return a list of popular HuggingFace models for suggestions"""
     try:
         from huggingface_hub import list_models
-        
+
         # Get popular text-generation models, sorted by downloads
         models = list_models(
             filter="text-generation",
             sort="downloads",
             direction=-1,
-            limit=20  # Get top 20 most downloaded
+            limit=20,  # Get top 20 most downloaded
         )
-        
+
         # Extract model names and filter for chat/conversation models
         model_names = []
-        chat_keywords = ['chat', 'instruct', 'dialog', 'conversation', 'assistant']
-        
+        chat_keywords = ["chat", "instruct", "dialog", "conversation", "assistant"]
+
         for model in models:
-            model_name = model.id if hasattr(model, 'id') else str(model)
+            model_name = model.id if hasattr(model, "id") else str(model)
             # Prioritize models with chat-related keywords
             if any(keyword in model_name.lower() for keyword in chat_keywords):
                 model_names.append(model_name)
             elif len(model_names) < 10:  # Fill up with other popular models
                 model_names.append(model_name)
-                
+
         return model_names[:10] if model_names else _get_fallback_hf_models()
-        
+
     except Exception:
         # Fallback to static list if API call fails
         return _get_fallback_hf_models()
 
 
-def _get_fallback_hf_models() -> List[str]:
+def _get_fallback_hf_models() -> list[str]:
     """Fallback list of popular HuggingFace models"""
     return [
         "microsoft/DialoGPT-medium",
-        "microsoft/DialoGPT-large", 
+        "microsoft/DialoGPT-large",
         "facebook/blenderbot-400M-distill",
         "microsoft/phi-2",
         "deepseek-ai/deepseek-llm-7b-chat",
@@ -165,44 +234,44 @@ def _get_fallback_hf_models() -> List[str]:
         "facebook/blenderbot_small-90M",
         "microsoft/phi-1_5",
         "facebook/opt-350m",
-        "EleutherAI/gpt-neo-1.3B"
+        "EleutherAI/gpt-neo-1.3B",
     ]
 
 
-def search_hf_models_fuzzy(query: str, limit: int = 10) -> List[str]:
+def search_hf_models_fuzzy(query: str, limit: int = 10) -> list[str]:
     """Use HuggingFace Hub's native fuzzy search for model suggestions"""
     try:
         from huggingface_hub import list_models
-        
+
         # HF Hub's search is already fuzzy! It handles typos and partial matches
         models = list_models(
             search=query,
             filter="text-generation",
-            sort="downloads", 
+            sort="downloads",
             direction=-1,
-            limit=limit
+            limit=limit,
         )
-        
-        model_names = [model.id if hasattr(model, 'id') else str(model) for model in models]
-        
+
+        model_names = [model.id if hasattr(model, "id") else str(model) for model in models]
+
         # If direct search doesn't return enough results, try some variations
         if len(model_names) < 3:
             # Try searching for partial matches or common variations
             variations = []
-            
+
             # Extract base name (e.g., "gpt3" from "gpt-3.5")
-            base_query = query.lower().replace('-', '').replace('.', '').replace('_', '')
+            base_query = query.lower().replace("-", "").replace(".", "").replace("_", "")
             if base_query != query.lower():
                 variations.append(base_query)
-            
+
             # Try common model name patterns
-            if 'gpt' in query.lower():
-                variations.extend(['gpt2', 'gpt-neo', 'gpt-j', 'dialoGPT'])
-            elif 'llama' in query.lower():
-                variations.extend(['llama2', 'alpaca', 'vicuna'])
-            elif 'bert' in query.lower():
-                variations.extend(['roberta', 'distilbert', 'albert'])
-            
+            if "gpt" in query.lower():
+                variations.extend(["gpt2", "gpt-neo", "gpt-j", "dialoGPT"])
+            elif "llama" in query.lower():
+                variations.extend(["llama2", "alpaca", "vicuna"])
+            elif "bert" in query.lower():
+                variations.extend(["roberta", "distilbert", "albert"])
+
             # Search with variations
             for var in variations[:2]:  # Limit to 2 variations to avoid too many API calls
                 try:
@@ -211,13 +280,15 @@ def search_hf_models_fuzzy(query: str, limit: int = 10) -> List[str]:
                         filter="text-generation",
                         sort="downloads",
                         direction=-1,
-                        limit=3
+                        limit=3,
                     )
-                    var_names = [model.id if hasattr(model, 'id') else str(model) for model in var_models]
+                    var_names = [
+                        model.id if hasattr(model, "id") else str(model) for model in var_models
+                    ]
                     model_names.extend(var_names)
-                except:
+                except Exception:
                     continue
-        
+
         # Remove duplicates while preserving order
         seen = set()
         unique_models = []
@@ -225,50 +296,96 @@ def search_hf_models_fuzzy(query: str, limit: int = 10) -> List[str]:
             if model not in seen:
                 seen.add(model)
                 unique_models.append(model)
-        
+
         return unique_models[:limit]
-        
+
     except Exception:
         # If search fails, return empty list
         return []
 
 
-def search_hf_models(query: str, limit: int = 10) -> List[str]:
+def search_hf_models(query: str, limit: int = 10) -> list[str]:
     """Simple search for HuggingFace models based on query (kept for backward compatibility)"""
     return search_hf_models_fuzzy(query, limit)
 
 
-def validate_model_and_suggest(model_name: str, llm_type: str) -> Optional[str]:
+def validate_model_and_suggest(model_name: str, llm_type: str) -> str | None:
     """Validate model name and provide suggestions if invalid"""
     if llm_type == "ollama":
         available_models = check_ollama_models()
         if available_models and model_name not in available_models:
-            # Use intelligent fuzzy search based on locally installed models
-            suggestions = search_ollama_models_fuzzy(model_name, available_models)
-            
             error_msg = f"Model '{model_name}' not found in your local Ollama installation."
-            if suggestions:
-                error_msg += "\n\nDid you mean one of these installed models?\n"
-                for i, suggestion in enumerate(suggestions, 1):
-                    error_msg += f"  {i}. {suggestion}\n"
+
+            # Check if the model exists remotely and get available tags
+            model_exists_remotely, available_tags = check_ollama_model_exists_remotely(model_name)
+
+            if model_exists_remotely and model_name in available_tags:
+                # Exact model exists remotely - suggest pulling it
+                error_msg += "\n\nTo install the requested model:\n"
+                error_msg += f"  ollama pull {model_name}\n"
+
+                # Show local alternatives
+                suggestions = search_ollama_models_fuzzy(model_name, available_models)
+                if suggestions:
+                    error_msg += "\nOr use one of these similar installed models:\n"
+                    for i, suggestion in enumerate(suggestions, 1):
+                        error_msg += f"  {i}. {suggestion}\n"
+
+            elif model_exists_remotely and available_tags:
+                # Base model exists but requested tag doesn't - suggest correct tags
+                base_model = model_name.split(":")[0]
+                requested_tag = model_name.split(":", 1)[1] if ":" in model_name else None
+
+                error_msg += (
+                    f"\n\nModel '{base_model}' exists, but tag '{requested_tag}' is not available."
+                )
+                error_msg += f"\n\nAvailable {base_model} models you can install:\n"
+                for i, tag in enumerate(available_tags[:8], 1):
+                    error_msg += f"  {i}. ollama pull {tag}\n"
+                if len(available_tags) > 8:
+                    error_msg += f"  ... and {len(available_tags) - 8} more variants\n"
+
+                # Also show local alternatives
+                suggestions = search_ollama_models_fuzzy(model_name, available_models)
+                if suggestions:
+                    error_msg += "\nOr use one of these similar installed models:\n"
+                    for i, suggestion in enumerate(suggestions, 1):
+                        error_msg += f"  {i}. {suggestion}\n"
+
             else:
-                error_msg += "\n\nYour installed models:\n"
-                for i, model in enumerate(available_models[:8], 1):
-                    error_msg += f"  {i}. {model}\n"
-                if len(available_models) > 8:
-                    error_msg += f"  ... and {len(available_models) - 8} more\n"
-            
-            error_msg += "\nTo list all models: ollama list"
-            error_msg += "\nTo download a new model: ollama pull <model_name>"
-            error_msg += "\nBrowse models: https://ollama.com/library"
+                # Model doesn't exist remotely - show fuzzy suggestions
+                suggestions = search_ollama_models_fuzzy(model_name, available_models)
+                error_msg += f"\n\nModel '{model_name}' was not found in Ollama's library."
+
+                if suggestions:
+                    error_msg += "\n\nDid you mean one of these installed models?\n"
+                    for i, suggestion in enumerate(suggestions, 1):
+                        error_msg += f"  {i}. {suggestion}\n"
+                else:
+                    error_msg += "\n\nYour installed models:\n"
+                    for i, model in enumerate(available_models[:8], 1):
+                        error_msg += f"  {i}. {model}\n"
+                    if len(available_models) > 8:
+                        error_msg += f"  ... and {len(available_models) - 8} more\n"
+
+            error_msg += "\n\nCommands:"
+            error_msg += "\n  ollama list                    # List installed models"
+            if model_exists_remotely and available_tags:
+                if model_name in available_tags:
+                    error_msg += f"\n  ollama pull {model_name}          # Install requested model"
+                else:
+                    error_msg += (
+                        f"\n  ollama pull {available_tags[0]}    # Install recommended variant"
+                    )
+            error_msg += "\n  https://ollama.com/library     # Browse available models"
             return error_msg
-            
+
     elif llm_type == "hf":
         # For HF models, we can do a quick existence check
         if not check_hf_model_exists(model_name):
             # Use HF Hub's native fuzzy search directly
             search_suggestions = search_hf_models_fuzzy(model_name, limit=8)
-            
+
             error_msg = f"Model '{model_name}' not found on HuggingFace Hub."
             if search_suggestions:
                 error_msg += "\n\nDid you mean one of these?\n"
@@ -280,10 +397,10 @@ def validate_model_and_suggest(model_name: str, llm_type: str) -> Optional[str]:
                 error_msg += "\n\nPopular chat models:\n"
                 for i, model in enumerate(popular_models[:5], 1):
                     error_msg += f"  {i}. {model}\n"
-            
+
             error_msg += f"\nSearch more: https://huggingface.co/models?search={model_name}&pipeline_tag=text-generation"
             return error_msg
-    
+
     return None  # Model is valid or we can't check
 
 
@@ -346,28 +463,27 @@ class OllamaChat(LLMInterface):
             # Check if the Ollama server is responsive
             if host:
                 requests.get(host)
-                
+
             # Pre-check model availability with helpful suggestions
             model_error = validate_model_and_suggest(model, "ollama")
             if model_error:
                 raise ValueError(model_error)
-                
+
         except ImportError:
             raise ImportError(
                 "The 'requests' library is required for Ollama. Please install it with 'pip install requests'."
             )
         except requests.exceptions.ConnectionError:
-            logger.error(
-                f"Could not connect to Ollama at {host}. Please ensure Ollama is running."
-            )
+            logger.error(f"Could not connect to Ollama at {host}. Please ensure Ollama is running.")
             raise ConnectionError(
                 f"Could not connect to Ollama at {host}. Please ensure Ollama is running."
             )
 
     def ask(self, prompt: str, **kwargs) -> str:
-        import requests
         import json
 
+        import requests
+
         full_url = f"{self.host}/api/generate"
         payload = {
             "model": self.model,
@@ -375,8 +491,9 @@ class OllamaChat(LLMInterface):
             "stream": False,  # Keep it simple for now
             "options": kwargs,
         }
-        logger.info(f"Sending request to Ollama: {payload}")
+        logger.debug(f"Sending request to Ollama: {payload}")
         try:
+            logger.info("Sending request to Ollama and waiting for response...")
             response = requests.post(full_url, data=json.dumps(payload))
             response.raise_for_status()
 
@@ -396,19 +513,19 @@ class OllamaChat(LLMInterface):
 
 
 class HFChat(LLMInterface):
-    """LLM interface for local Hugging Face Transformers models."""
+    """LLM interface for local Hugging Face Transformers models with proper chat templates."""
 
     def __init__(self, model_name: str = "deepseek-ai/deepseek-llm-7b-chat"):
         logger.info(f"Initializing HFChat with model='{model_name}'")
-        
+
         # Pre-check model availability with helpful suggestions
         model_error = validate_model_and_suggest(model_name, "hf")
         if model_error:
             raise ValueError(model_error)
-            
+
         try:
-            from transformers.pipelines import pipeline
             import torch
+            from transformers import AutoModelForCausalLM, AutoTokenizer
         except ImportError:
             raise ImportError(
                 "The 'transformers' and 'torch' libraries are required for Hugging Face models. Please install them with 'pip install transformers torch'."
@@ -416,60 +533,129 @@ class HFChat(LLMInterface):
 
         # Auto-detect device
         if torch.cuda.is_available():
-            device = "cuda"
+            self.device = "cuda"
             logger.info("CUDA is available. Using GPU.")
         elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
-            device = "mps"
+            self.device = "mps"
             logger.info("MPS is available. Using Apple Silicon GPU.")
         else:
-            device = "cpu"
+            self.device = "cpu"
             logger.info("No GPU detected. Using CPU.")
 
-        self.pipeline = pipeline("text-generation", model=model_name, device=device)
+        # Load tokenizer and model with timeout protection
+        try:
+            import signal
+
+            def timeout_handler(signum, frame):
+                raise TimeoutError("Model download/loading timed out")
+
+            # Set timeout for model loading (60 seconds)
+            old_handler = signal.signal(signal.SIGALRM, timeout_handler)
+            signal.alarm(60)
+
+            try:
+                logger.info(f"Loading tokenizer for {model_name}...")
+                self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+                logger.info(f"Loading model {model_name}...")
+                self.model = AutoModelForCausalLM.from_pretrained(
+                    model_name,
+                    torch_dtype=torch.float16 if self.device != "cpu" else torch.float32,
+                    device_map="auto" if self.device != "cpu" else None,
+                    trust_remote_code=True,
+                )
+                logger.info(f"Successfully loaded {model_name}")
+            finally:
+                signal.alarm(0)  # Cancel the alarm
+                signal.signal(signal.SIGALRM, old_handler)  # Restore old handler
+
+        except TimeoutError:
+            logger.error(f"Model loading timed out for {model_name}")
+            raise RuntimeError(
+                f"Model loading timed out for {model_name}. Please check your internet connection or try a smaller model."
+            )
+        except Exception as e:
+            logger.error(f"Failed to load model {model_name}: {e}")
+            raise
+
+        # Move model to device if not using device_map
+        if self.device != "cpu" and "device_map" not in str(self.model):
+            self.model = self.model.to(self.device)
+
+        # Set pad token if not present
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
 
     def ask(self, prompt: str, **kwargs) -> str:
-        # Map OpenAI-style arguments to Hugging Face equivalents
-        if "max_tokens" in kwargs:
-            # Prefer user-provided max_new_tokens if both are present
-            kwargs.setdefault("max_new_tokens", kwargs["max_tokens"])
-            # Remove the unsupported key to avoid errors in Transformers
-            kwargs.pop("max_tokens")
+        print("kwargs in HF: ", kwargs)
+        # Check if this is a Qwen model and add /no_think by default
+        is_qwen_model = "qwen" in self.model.config._name_or_path.lower()
 
-        # Handle temperature=0 edge-case for greedy decoding
-        if "temperature" in kwargs and kwargs["temperature"] == 0.0:
-            # Remove unsupported zero temperature and use deterministic generation
-            kwargs.pop("temperature")
-            kwargs.setdefault("do_sample", False)
+        # For Qwen models, automatically add /no_think to the prompt
+        if is_qwen_model and "/no_think" not in prompt and "/think" not in prompt:
+            prompt = prompt + " /no_think"
 
-        # Sensible defaults for text generation
-        params = {"max_length": 500, "num_return_sequences": 1, **kwargs}
-        logger.info(f"Generating text with Hugging Face model with params: {params}")
-        results = self.pipeline(prompt, **params)
+        # Prepare chat template
+        messages = [{"role": "user", "content": prompt}]
 
-        # Handle different response formats from transformers
-        if isinstance(results, list) and len(results) > 0:
-            generated_text = (
-                results[0].get("generated_text", "")
-                if isinstance(results[0], dict)
-                else str(results[0])
-            )
+        # Apply chat template if available
+        if hasattr(self.tokenizer, "apply_chat_template"):
+            try:
+                formatted_prompt = self.tokenizer.apply_chat_template(
+                    messages, tokenize=False, add_generation_prompt=True
+                )
+            except Exception as e:
+                logger.warning(f"Chat template failed, using raw prompt: {e}")
+                formatted_prompt = prompt
         else:
-            generated_text = str(results)
+            # Fallback for models without chat template
+            formatted_prompt = prompt
 
-        # Extract only the newly generated portion by removing the original prompt
-        if isinstance(generated_text, str) and generated_text.startswith(prompt):
-            response = generated_text[len(prompt) :].strip()
-        else:
-            # Fallback: return the full response if prompt removal fails
-            response = str(generated_text)
+        # Tokenize input
+        inputs = self.tokenizer(
+            formatted_prompt,
+            return_tensors="pt",
+            padding=True,
+            truncation=True,
+            max_length=2048,
+        )
 
-        return response
+        # Move inputs to device
+        if self.device != "cpu":
+            inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+        # Set generation parameters
+        generation_config = {
+            "max_new_tokens": kwargs.get("max_tokens", kwargs.get("max_new_tokens", 512)),
+            "temperature": kwargs.get("temperature", 0.7),
+            "top_p": kwargs.get("top_p", 0.9),
+            "do_sample": kwargs.get("temperature", 0.7) > 0,
+            "pad_token_id": self.tokenizer.eos_token_id,
+            "eos_token_id": self.tokenizer.eos_token_id,
+        }
+
+        # Handle temperature=0 for greedy decoding
+        if generation_config["temperature"] == 0.0:
+            generation_config["do_sample"] = False
+            generation_config.pop("temperature")
+
+        logger.info(f"Generating with HuggingFace model, config: {generation_config}")
+
+        # Generate
+        with torch.no_grad():
+            outputs = self.model.generate(**inputs, **generation_config)
+
+        # Decode response
+        generated_tokens = outputs[0][inputs["input_ids"].shape[1] :]
+        response = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
+
+        return response.strip()
 
 
 class OpenAIChat(LLMInterface):
     """LLM interface for OpenAI models."""
 
-    def __init__(self, model: str = "gpt-4o", api_key: Optional[str] = None):
+    def __init__(self, model: str = "gpt-4o", api_key: str | None = None):
         self.model = model
         self.api_key = api_key or os.getenv("OPENAI_API_KEY")
 
@@ -496,11 +682,7 @@ class OpenAIChat(LLMInterface):
             "messages": [{"role": "user", "content": prompt}],
             "max_tokens": kwargs.get("max_tokens", 1000),
             "temperature": kwargs.get("temperature", 0.7),
-            **{
-                k: v
-                for k, v in kwargs.items()
-                if k not in ["max_tokens", "temperature"]
-            },
+            **{k: v for k, v in kwargs.items() if k not in ["max_tokens", "temperature"]},
         }
 
         logger.info(f"Sending request to OpenAI with model {self.model}")
@@ -522,7 +704,7 @@ class SimulatedChat(LLMInterface):
         return "This is a simulated answer from the LLM based on the retrieved context."
 
 
-def get_llm(llm_config: Optional[Dict[str, Any]] = None) -> LLMInterface:
+def get_llm(llm_config: dict[str, Any] | None = None) -> LLMInterface:
     """
     Factory function to get an LLM interface based on configuration.
 

packages/leann-core/src/leann/cli.py

@@ -1,35 +1,62 @@
-#!/usr/bin/env python3
 import argparse
 import asyncio
-import sys
 from pathlib import Path
-from typing import Optional
-import os
 
 from llama_index.core import SimpleDirectoryReader
 from llama_index.core.node_parser import SentenceSplitter
 
-from .api import LeannBuilder, LeannSearcher, LeannChat
+from .api import LeannBuilder, LeannChat, LeannSearcher
+
+
+def extract_pdf_text_with_pymupdf(file_path: str) -> str:
+    """Extract text from PDF using PyMuPDF for better quality."""
+    try:
+        import fitz  # PyMuPDF
+
+        doc = fitz.open(file_path)
+        text = ""
+        for page in doc:
+            text += page.get_text()
+        doc.close()
+        return text
+    except ImportError:
+        # Fallback to default reader
+        return None
+
+
+def extract_pdf_text_with_pdfplumber(file_path: str) -> str:
+    """Extract text from PDF using pdfplumber for better quality."""
+    try:
+        import pdfplumber
+
+        text = ""
+        with pdfplumber.open(file_path) as pdf:
+            for page in pdf.pages:
+                text += page.extract_text() or ""
+        return text
+    except ImportError:
+        # Fallback to default reader
+        return None
 
 
 class LeannCLI:
     def __init__(self):
         self.indexes_dir = Path.home() / ".leann" / "indexes"
         self.indexes_dir.mkdir(parents=True, exist_ok=True)
-        
+
         self.node_parser = SentenceSplitter(
             chunk_size=256, chunk_overlap=128, separator=" ", paragraph_separator="\n\n"
         )
-    
+
     def get_index_path(self, index_name: str) -> str:
         index_dir = self.indexes_dir / index_name
         return str(index_dir / "documents.leann")
-    
+
     def index_exists(self, index_name: str) -> bool:
         index_dir = self.indexes_dir / index_name
         meta_file = index_dir / "documents.leann.meta.json"
         return meta_file.exists()
-    
+
     def create_parser(self) -> argparse.ArgumentParser:
         parser = argparse.ArgumentParser(
             prog="leann",
@@ -41,16 +68,18 @@ Examples:
   leann search my-docs "query"             # Search in my-docs index
   leann ask my-docs "question"             # Ask my-docs index
   leann list                              # List all stored indexes
-            """
+            """,
         )
-        
+
         subparsers = parser.add_subparsers(dest="command", help="Available commands")
-        
+
         # Build command
         build_parser = subparsers.add_parser("build", help="Build document index")
         build_parser.add_argument("index_name", help="Index name")
         build_parser.add_argument("--docs", type=str, required=True, help="Documents directory")
-        build_parser.add_argument("--backend", type=str, default="hnsw", choices=["hnsw", "diskann"])
+        build_parser.add_argument(
+            "--backend", type=str, default="hnsw", choices=["hnsw", "diskann"]
+        )
         build_parser.add_argument("--embedding-model", type=str, default="facebook/contriever")
         build_parser.add_argument("--force", "-f", action="store_true", help="Force rebuild")
         build_parser.add_argument("--graph-degree", type=int, default=32)
@@ -58,7 +87,7 @@ Examples:
         build_parser.add_argument("--num-threads", type=int, default=1)
         build_parser.add_argument("--compact", action="store_true", default=True)
         build_parser.add_argument("--recompute", action="store_true", default=True)
-        
+
         # Search command
         search_parser = subparsers.add_parser("search", help="Search documents")
         search_parser.add_argument("index_name", help="Index name")
@@ -68,12 +97,21 @@ Examples:
         search_parser.add_argument("--beam-width", type=int, default=1)
         search_parser.add_argument("--prune-ratio", type=float, default=0.0)
         search_parser.add_argument("--recompute-embeddings", action="store_true")
-        search_parser.add_argument("--pruning-strategy", choices=["global", "local", "proportional"], default="global")
-        
+        search_parser.add_argument(
+            "--pruning-strategy",
+            choices=["global", "local", "proportional"],
+            default="global",
+        )
+
         # Ask command
         ask_parser = subparsers.add_parser("ask", help="Ask questions")
         ask_parser.add_argument("index_name", help="Index name")
-        ask_parser.add_argument("--llm", type=str, default="ollama", choices=["simulated", "ollama", "hf", "openai"])
+        ask_parser.add_argument(
+            "--llm",
+            type=str,
+            default="ollama",
+            choices=["simulated", "ollama", "hf", "openai"],
+        )
         ask_parser.add_argument("--model", type=str, default="qwen3:8b")
         ask_parser.add_argument("--host", type=str, default="http://localhost:11434")
         ask_parser.add_argument("--interactive", "-i", action="store_true")
@@ -82,81 +120,118 @@ Examples:
         ask_parser.add_argument("--beam-width", type=int, default=1)
         ask_parser.add_argument("--prune-ratio", type=float, default=0.0)
         ask_parser.add_argument("--recompute-embeddings", action="store_true")
-        ask_parser.add_argument("--pruning-strategy", choices=["global", "local", "proportional"], default="global")
-        
+        ask_parser.add_argument(
+            "--pruning-strategy",
+            choices=["global", "local", "proportional"],
+            default="global",
+        )
+
         # List command
-        list_parser = subparsers.add_parser("list", help="List all indexes")
-        
+        subparsers.add_parser("list", help="List all indexes")
+
         return parser
-    
+
     def list_indexes(self):
         print("Stored LEANN indexes:")
-        
+
         if not self.indexes_dir.exists():
             print("No indexes found. Use 'leann build <name> --docs <dir>' to create one.")
             return
-        
+
         index_dirs = [d for d in self.indexes_dir.iterdir() if d.is_dir()]
-        
+
         if not index_dirs:
             print("No indexes found. Use 'leann build <name> --docs <dir>' to create one.")
             return
-        
+
         print(f"Found {len(index_dirs)} indexes:")
         for i, index_dir in enumerate(index_dirs, 1):
             index_name = index_dir.name
             status = "✓" if self.index_exists(index_name) else "✗"
-            
+
             print(f"  {i}. {index_name} [{status}]")
             if self.index_exists(index_name):
-                meta_file = index_dir / "documents.leann.meta.json"
-                size_mb = sum(f.stat().st_size for f in index_dir.iterdir() if f.is_file()) / (1024 * 1024)
+                index_dir / "documents.leann.meta.json"
+                size_mb = sum(f.stat().st_size for f in index_dir.iterdir() if f.is_file()) / (
+                    1024 * 1024
+                )
                 print(f"     Size: {size_mb:.1f} MB")
-        
+
         if index_dirs:
             example_name = index_dirs[0].name
-            print(f"\nUsage:")
-            print(f"  leann search {example_name} \"your query\"")
+            print("\nUsage:")
+            print(f'  leann search {example_name} "your query"')
             print(f"  leann ask {example_name} --interactive")
-    
+
     def load_documents(self, docs_dir: str):
         print(f"Loading documents from {docs_dir}...")
-        
-        documents = SimpleDirectoryReader(
+
+        # Try to use better PDF parsers first
+        documents = []
+        docs_path = Path(docs_dir)
+
+        for file_path in docs_path.rglob("*.pdf"):
+            print(f"Processing PDF: {file_path}")
+
+            # Try PyMuPDF first (best quality)
+            text = extract_pdf_text_with_pymupdf(str(file_path))
+            if text is None:
+                # Try pdfplumber
+                text = extract_pdf_text_with_pdfplumber(str(file_path))
+
+            if text:
+                # Create a simple document structure
+                from llama_index.core import Document
+
+                doc = Document(text=text, metadata={"source": str(file_path)})
+                documents.append(doc)
+            else:
+                # Fallback to default reader
+                print(f"Using default reader for {file_path}")
+                default_docs = SimpleDirectoryReader(
+                    str(file_path.parent),
+                    filename_as_id=True,
+                    required_exts=[file_path.suffix],
+                ).load_data()
+                documents.extend(default_docs)
+
+        # Load other file types with default reader
+        other_docs = SimpleDirectoryReader(
             docs_dir,
             recursive=True,
             encoding="utf-8",
-            required_exts=[".pdf", ".txt", ".md", ".docx"],
+            required_exts=[".txt", ".md", ".docx"],
         ).load_data(show_progress=True)
-        
+        documents.extend(other_docs)
+
         all_texts = []
         for doc in documents:
             nodes = self.node_parser.get_nodes_from_documents([doc])
             for node in nodes:
                 all_texts.append(node.get_content())
-        
+
         print(f"Loaded {len(documents)} documents, {len(all_texts)} chunks")
         return all_texts
-    
+
     async def build_index(self, args):
         docs_dir = args.docs
         index_name = args.index_name
         index_dir = self.indexes_dir / index_name
         index_path = self.get_index_path(index_name)
-        
+
         if index_dir.exists() and not args.force:
             print(f"Index '{index_name}' already exists. Use --force to rebuild.")
             return
-        
+
         all_texts = self.load_documents(docs_dir)
         if not all_texts:
             print("No documents found")
             return
-        
+
         index_dir.mkdir(parents=True, exist_ok=True)
-        
+
         print(f"Building index '{index_name}' with {args.backend} backend...")
-        
+
         builder = LeannBuilder(
             backend_name=args.backend,
             embedding_model=args.embedding_model,
@@ -166,103 +241,107 @@ Examples:
             is_recompute=args.recompute,
             num_threads=args.num_threads,
         )
-        
+
         for chunk_text in all_texts:
             builder.add_text(chunk_text)
-        
+
         builder.build_index(index_path)
         print(f"Index built at {index_path}")
-    
+
     async def search_documents(self, args):
         index_name = args.index_name
         query = args.query
         index_path = self.get_index_path(index_name)
-        
+
         if not self.index_exists(index_name):
-            print(f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir>' to create it.")
+            print(
+                f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir>' to create it."
+            )
             return
-        
+
         searcher = LeannSearcher(index_path=index_path)
         results = searcher.search(
-            query, 
+            query,
             top_k=args.top_k,
             complexity=args.complexity,
             beam_width=args.beam_width,
             prune_ratio=args.prune_ratio,
             recompute_embeddings=args.recompute_embeddings,
-            pruning_strategy=args.pruning_strategy
+            pruning_strategy=args.pruning_strategy,
         )
-        
+
         print(f"Search results for '{query}' (top {len(results)}):")
         for i, result in enumerate(results, 1):
             print(f"{i}. Score: {result.score:.3f}")
             print(f"   {result.text[:200]}...")
             print()
-    
+
     async def ask_questions(self, args):
         index_name = args.index_name
         index_path = self.get_index_path(index_name)
-        
+
         if not self.index_exists(index_name):
-            print(f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir>' to create it.")
+            print(
+                f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir>' to create it."
+            )
             return
-        
+
         print(f"Starting chat with index '{index_name}'...")
         print(f"Using {args.model} ({args.llm})")
-        
+
         llm_config = {"type": args.llm, "model": args.model}
         if args.llm == "ollama":
             llm_config["host"] = args.host
-        
+
         chat = LeannChat(index_path=index_path, llm_config=llm_config)
-        
+
         if args.interactive:
             print("LEANN Assistant ready! Type 'quit' to exit")
             print("=" * 40)
-            
+
             while True:
                 user_input = input("\nYou: ").strip()
-                if user_input.lower() in ['quit', 'exit', 'q']:
+                if user_input.lower() in ["quit", "exit", "q"]:
                     print("Goodbye!")
                     break
-                
+
                 if not user_input:
                     continue
-                
+
                 response = chat.ask(
-                    user_input, 
+                    user_input,
                     top_k=args.top_k,
                     complexity=args.complexity,
                     beam_width=args.beam_width,
                     prune_ratio=args.prune_ratio,
                     recompute_embeddings=args.recompute_embeddings,
-                    pruning_strategy=args.pruning_strategy
+                    pruning_strategy=args.pruning_strategy,
                 )
                 print(f"LEANN: {response}")
         else:
             query = input("Enter your question: ").strip()
             if query:
                 response = chat.ask(
-                    query, 
+                    query,
                     top_k=args.top_k,
                     complexity=args.complexity,
                     beam_width=args.beam_width,
                     prune_ratio=args.prune_ratio,
                     recompute_embeddings=args.recompute_embeddings,
-                    pruning_strategy=args.pruning_strategy
+                    pruning_strategy=args.pruning_strategy,
                 )
                 print(f"LEANN: {response}")
-    
+
     async def run(self, args=None):
         parser = self.create_parser()
-        
+
         if args is None:
             args = parser.parse_args()
-        
+
         if not args.command:
             parser.print_help()
             return
-        
+
         if args.command == "list":
             self.list_indexes()
         elif args.command == "build":
@@ -277,11 +356,12 @@ Examples:
 
 def main():
     import dotenv
+
     dotenv.load_dotenv()
-    
+
     cli = LeannCLI()
     asyncio.run(cli.run())
 
 
 if __name__ == "__main__":
-    main()
+    main()

packages/leann-core/src/leann/embedding_compute.py

@@ -4,16 +4,30 @@ Consolidates all embedding computation logic using SentenceTransformer
 Preserves all optimization parameters to ensure performance
 """
 
+import logging
+import os
+from typing import Any
+
 import numpy as np
 import torch
-from typing import List
-import logging
 
+# Set up logger with proper level
 logger = logging.getLogger(__name__)
+LOG_LEVEL = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
+log_level = getattr(logging, LOG_LEVEL, logging.WARNING)
+logger.setLevel(log_level)
+
+# Global model cache to avoid repeated loading
+_model_cache: dict[str, Any] = {}
 
 
 def compute_embeddings(
-    texts: List[str], model_name: str, mode: str = "sentence-transformers"
+    texts: list[str],
+    model_name: str,
+    mode: str = "sentence-transformers",
+    is_build: bool = False,
+    batch_size: int = 32,
+    adaptive_optimization: bool = True,
 ) -> np.ndarray:
     """
     Unified embedding computation entry point
@@ -22,12 +36,21 @@ def compute_embeddings(
         texts: List of texts to compute embeddings for
         model_name: Model name
         mode: Computation mode ('sentence-transformers', 'openai', 'mlx')
+        is_build: Whether this is a build operation (shows progress bar)
+        batch_size: Batch size for processing
+        adaptive_optimization: Whether to use adaptive optimization based on batch size
 
     Returns:
         Normalized embeddings array, shape: (len(texts), embedding_dim)
     """
     if mode == "sentence-transformers":
-        return compute_embeddings_sentence_transformers(texts, model_name)
+        return compute_embeddings_sentence_transformers(
+            texts,
+            model_name,
+            is_build=is_build,
+            batch_size=batch_size,
+            adaptive_optimization=adaptive_optimization,
+        )
     elif mode == "openai":
         return compute_embeddings_openai(texts, model_name)
     elif mode == "mlx":
@@ -37,32 +60,33 @@ def compute_embeddings(
 
 
 def compute_embeddings_sentence_transformers(
-    texts: List[str],
+    texts: list[str],
     model_name: str,
     use_fp16: bool = True,
     device: str = "auto",
     batch_size: int = 32,
+    is_build: bool = False,
+    adaptive_optimization: bool = True,
 ) -> np.ndarray:
     """
-    Compute embeddings using SentenceTransformer
-    Preserves all optimization parameters to ensure consistency with original embedding_server
+    Compute embeddings using SentenceTransformer with model caching and adaptive optimization
 
     Args:
         texts: List of texts to compute embeddings for
-        model_name: SentenceTransformer model name
+        model_name: Model name
         use_fp16: Whether to use FP16 precision
-        device: Device selection ('auto', 'cuda', 'mps', 'cpu')
+        device: Device to use ('auto', 'cuda', 'mps', 'cpu')
         batch_size: Batch size for processing
-
-    Returns:
-        Normalized embeddings array, shape: (len(texts), embedding_dim)
+        is_build: Whether this is a build operation (shows progress bar)
+        adaptive_optimization: Whether to use adaptive optimization based on batch size
     """
-    print(
-        f"INFO: Computing embeddings for {len(texts)} texts using SentenceTransformer, model: '{model_name}'"
+    # Handle empty input
+    if not texts:
+        raise ValueError("Cannot compute embeddings for empty text list")
+    logger.info(
+        f"Computing embeddings for {len(texts)} texts using SentenceTransformer, model: '{model_name}'"
     )
 
-    from sentence_transformers import SentenceTransformer
-
     # Auto-detect device
     if device == "auto":
         if torch.cuda.is_available():
@@ -72,91 +96,149 @@ def compute_embeddings_sentence_transformers(
         else:
             device = "cpu"
 
-    print(f"INFO: Using device: {device}")
+    # Apply optimizations based on benchmark results
+    if adaptive_optimization:
+        # Use optimal batch_size constants for different devices based on benchmark results
+        if device == "mps":
+            batch_size = 128  # MPS optimal batch size from benchmark
+            if model_name == "Qwen/Qwen3-Embedding-0.6B":
+                batch_size = 32
+        elif device == "cuda":
+            batch_size = 256  # CUDA optimal batch size
+        # Keep original batch_size for CPU
 
-    # Prepare model and tokenizer optimization parameters (consistent with original embedding_server)
-    model_kwargs = {
-        "torch_dtype": torch.float16 if use_fp16 else torch.float32,
-        "low_cpu_mem_usage": True,
-        "_fast_init": True,  # Skip weight initialization checks for faster loading
-    }
+    # Create cache key
+    cache_key = f"sentence_transformers_{model_name}_{device}_{use_fp16}_optimized"
 
-    tokenizer_kwargs = {
-        "use_fast": True,  # Use fast tokenizer for better runtime performance
-    }
+    # Check if model is already cached
+    if cache_key in _model_cache:
+        logger.info(f"Using cached optimized model: {model_name}")
+        model = _model_cache[cache_key]
+    else:
+        logger.info(f"Loading and caching optimized SentenceTransformer model: {model_name}")
+        from sentence_transformers import SentenceTransformer
 
-    # Load SentenceTransformer (try local first, then network)
-    print(f"INFO: Loading SentenceTransformer model: {model_name}")
+        logger.info(f"Using device: {device}")
 
-    try:
-        # Try local loading (avoid network delays)
-        model_kwargs["local_files_only"] = True
-        tokenizer_kwargs["local_files_only"] = True
+        # Apply hardware optimizations
+        if device == "cuda":
+            # TODO: Haven't tested this yet
+            torch.backends.cuda.matmul.allow_tf32 = True
+            torch.backends.cudnn.allow_tf32 = True
+            torch.backends.cudnn.benchmark = True
+            torch.backends.cudnn.deterministic = False
+            torch.cuda.set_per_process_memory_fraction(0.9)
+        elif device == "mps":
+            try:
+                if hasattr(torch.mps, "set_per_process_memory_fraction"):
+                    torch.mps.set_per_process_memory_fraction(0.9)
+            except AttributeError:
+                logger.warning("Some MPS optimizations not available in this PyTorch version")
+        elif device == "cpu":
+            # TODO: Haven't tested this yet
+            torch.set_num_threads(min(8, os.cpu_count() or 4))
+            try:
+                torch.backends.mkldnn.enabled = True
+            except AttributeError:
+                pass
 
-        model = SentenceTransformer(
-            model_name,
-            device=device,
-            model_kwargs=model_kwargs,
-            tokenizer_kwargs=tokenizer_kwargs,
-            local_files_only=True,
-        )
-        print("✅ Model loaded successfully! (local + optimized)")
-    except Exception as e:
-        print(f"Local loading failed ({e}), trying network download...")
-        # Fallback to network loading
-        model_kwargs["local_files_only"] = False
-        tokenizer_kwargs["local_files_only"] = False
+        # Prepare optimized model and tokenizer parameters
+        model_kwargs = {
+            "torch_dtype": torch.float16 if use_fp16 else torch.float32,
+            "low_cpu_mem_usage": True,
+            "_fast_init": True,
+            "attn_implementation": "eager",  # Use eager attention for speed
+        }
 
-        model = SentenceTransformer(
-            model_name,
-            device=device,
-            model_kwargs=model_kwargs,
-            tokenizer_kwargs=tokenizer_kwargs,
-            local_files_only=False,
-        )
-        print("✅ Model loaded successfully! (network + optimized)")
+        tokenizer_kwargs = {
+            "use_fast": True,
+            "padding": True,
+            "truncation": True,
+        }
 
-    # Apply additional optimizations (if supported)
-    if use_fp16 and device in ["cuda", "mps"]:
         try:
-            model = model.half()
-            model = torch.compile(model)
-            print(f"✅ Using FP16 precision and compile optimization: {model_name}")
-        except Exception as e:
-            print(
-                f"FP16 or compile optimization failed, continuing with default settings: {e}"
+            # Try local loading first
+            model_kwargs["local_files_only"] = True
+            tokenizer_kwargs["local_files_only"] = True
+
+            model = SentenceTransformer(
+                model_name,
+                device=device,
+                model_kwargs=model_kwargs,
+                tokenizer_kwargs=tokenizer_kwargs,
+                local_files_only=True,
             )
+            logger.info("Model loaded successfully! (local + optimized)")
+        except Exception as e:
+            logger.warning(f"Local loading failed ({e}), trying network download...")
+            # Fallback to network loading
+            model_kwargs["local_files_only"] = False
+            tokenizer_kwargs["local_files_only"] = False
 
-    # Compute embeddings (using SentenceTransformer's optimized implementation)
-    print("INFO: Starting embedding computation...")
+            model = SentenceTransformer(
+                model_name,
+                device=device,
+                model_kwargs=model_kwargs,
+                tokenizer_kwargs=tokenizer_kwargs,
+                local_files_only=False,
+            )
+            logger.info("Model loaded successfully! (network + optimized)")
 
-    embeddings = model.encode(
-        texts,
-        batch_size=batch_size,
-        show_progress_bar=False,  # Don't show progress bar in server environment
-        convert_to_numpy=True,
-        normalize_embeddings=False,  # Keep consistent with original API behavior
-        device=device,
-    )
+        # Apply additional optimizations based on mode
+        if use_fp16 and device in ["cuda", "mps"]:
+            try:
+                model = model.half()
+                logger.info(f"Applied FP16 precision: {model_name}")
+            except Exception as e:
+                logger.warning(f"FP16 optimization failed: {e}")
 
-    print(
-        f"INFO: Generated {len(embeddings)} embeddings, dimension: {embeddings.shape[1]}"
-    )
+        # Apply torch.compile optimization
+        if device in ["cuda", "mps"]:
+            try:
+                model = torch.compile(model, mode="reduce-overhead", dynamic=True)
+                logger.info(f"Applied torch.compile optimization: {model_name}")
+            except Exception as e:
+                logger.warning(f"torch.compile optimization failed: {e}")
+
+        # Set model to eval mode and disable gradients for inference
+        model.eval()
+        for param in model.parameters():
+            param.requires_grad_(False)
+
+        # Cache the model
+        _model_cache[cache_key] = model
+        logger.info(f"Model cached: {cache_key}")
+
+    # Compute embeddings with optimized inference mode
+    logger.info(f"Starting embedding computation... (batch_size: {batch_size})")
+
+    # Use torch.inference_mode for optimal performance
+    with torch.inference_mode():
+        embeddings = model.encode(
+            texts,
+            batch_size=batch_size,
+            show_progress_bar=is_build,  # Don't show progress bar in server environment
+            convert_to_numpy=True,
+            normalize_embeddings=False,
+            device=device,
+        )
+
+    logger.info(f"Generated {len(embeddings)} embeddings, dimension: {embeddings.shape[1]}")
 
     # Validate results
     if np.isnan(embeddings).any() or np.isinf(embeddings).any():
-        raise RuntimeError(
-            f"Detected NaN or Inf values in embeddings, model: {model_name}"
-        )
+        raise RuntimeError(f"Detected NaN or Inf values in embeddings, model: {model_name}")
 
     return embeddings
 
 
-def compute_embeddings_openai(texts: List[str], model_name: str) -> np.ndarray:
+def compute_embeddings_openai(texts: list[str], model_name: str) -> np.ndarray:
+    # TODO: @yichuan-w add progress bar only in build mode
     """Compute embeddings using OpenAI API"""
     try:
-        import openai
         import os
+
+        import openai
     except ImportError as e:
         raise ImportError(f"OpenAI package not installed: {e}")
 
@@ -164,14 +246,22 @@ def compute_embeddings_openai(texts: List[str], model_name: str) -> np.ndarray:
     if not api_key:
         raise RuntimeError("OPENAI_API_KEY environment variable not set")
 
-    client = openai.OpenAI(api_key=api_key)
+    # Cache OpenAI client
+    cache_key = "openai_client"
+    if cache_key in _model_cache:
+        client = _model_cache[cache_key]
+    else:
+        client = openai.OpenAI(api_key=api_key)
+        _model_cache[cache_key] = client
+        logger.info("OpenAI client cached")
 
-    print(
-        f"INFO: Computing embeddings for {len(texts)} texts using OpenAI API, model: '{model_name}'"
+    logger.info(
+        f"Computing embeddings for {len(texts)} texts using OpenAI API, model: '{model_name}'"
     )
+    print(f"len of texts: {len(texts)}")
 
     # OpenAI has limits on batch size and input length
-    max_batch_size = 100  # Conservative batch size
+    max_batch_size = 1000  # Conservative batch size
     all_embeddings = []
 
     try:
@@ -194,35 +284,40 @@ def compute_embeddings_openai(texts: List[str], model_name: str) -> np.ndarray:
             batch_embeddings = [embedding.embedding for embedding in response.data]
             all_embeddings.extend(batch_embeddings)
         except Exception as e:
-            print(f"ERROR: Batch {i} failed: {e}")
+            logger.error(f"Batch {i} failed: {e}")
             raise
 
     embeddings = np.array(all_embeddings, dtype=np.float32)
-    print(
-        f"INFO: Generated {len(embeddings)} embeddings, dimension: {embeddings.shape[1]}"
-    )
+    logger.info(f"Generated {len(embeddings)} embeddings, dimension: {embeddings.shape[1]}")
+    print(f"len of embeddings: {len(embeddings)}")
     return embeddings
 
 
-def compute_embeddings_mlx(
-    chunks: List[str], model_name: str, batch_size: int = 16
-) -> np.ndarray:
+def compute_embeddings_mlx(chunks: list[str], model_name: str, batch_size: int = 16) -> np.ndarray:
+    # TODO: @yichuan-w add progress bar only in build mode
     """Computes embeddings using an MLX model."""
     try:
         import mlx.core as mx
         from mlx_lm.utils import load
-        from tqdm import tqdm
     except ImportError as e:
         raise RuntimeError(
             "MLX or related libraries not available. Install with: uv pip install mlx mlx-lm"
         ) from e
 
-    print(
-        f"INFO: Computing embeddings for {len(chunks)} chunks using MLX model '{model_name}' with batch_size={batch_size}..."
+    logger.info(
+        f"Computing embeddings for {len(chunks)} chunks using MLX model '{model_name}' with batch_size={batch_size}..."
     )
 
-    # Load model and tokenizer
-    model, tokenizer = load(model_name)
+    # Cache MLX model and tokenizer
+    cache_key = f"mlx_{model_name}"
+    if cache_key in _model_cache:
+        logger.info(f"Using cached MLX model: {model_name}")
+        model, tokenizer = _model_cache[cache_key]
+    else:
+        logger.info(f"Loading and caching MLX model: {model_name}")
+        model, tokenizer = load(model_name)
+        _model_cache[cache_key] = (model, tokenizer)
+        logger.info(f"MLX model cached: {cache_key}")
 
     # Process chunks in batches with progress bar
     all_embeddings = []

packages/leann-core/src/leann/embedding_server_manager.py

@@ -1,14 +1,40 @@
-import threading
-import time
 import atexit
+import logging
+import os
 import socket
 import subprocess
 import sys
+import time
 from pathlib import Path
-from typing import Optional
-import select
+
 import psutil
 
+# Set up logging based on environment variable
+LOG_LEVEL = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
+logging.basicConfig(
+    level=getattr(logging, LOG_LEVEL, logging.INFO),
+    format="%(levelname)s - %(name)s - %(message)s",
+)
+logger = logging.getLogger(__name__)
+
+
+def _is_colab_environment() -> bool:
+    """Check if we're running in Google Colab environment."""
+    return "COLAB_GPU" in os.environ or "COLAB_TPU" in os.environ
+
+
+def _get_available_port(start_port: int = 5557) -> int:
+    """Get an available port starting from start_port."""
+    port = start_port
+    while port < start_port + 100:  # Try up to 100 ports
+        try:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+                s.bind(("localhost", port))
+                return port
+        except OSError:
+            port += 1
+    raise RuntimeError(f"No available ports found in range {start_port}-{start_port + 100}")
+
 
 def _check_port(port: int) -> bool:
     """Check if a port is in use"""
@@ -36,11 +62,11 @@ def _check_process_matches_config(
                 cmdline, port, expected_model, expected_passages_file
             )
 
-        print(f"DEBUG: No process found listening on port {port}")
+        logger.debug(f"No process found listening on port {port}")
         return False
 
     except Exception as e:
-        print(f"WARNING: Could not check process on port {port}: {e}")
+        logger.warning(f"Could not check process on port {port}: {e}")
         return False
 
 
@@ -61,7 +87,7 @@ def _check_cmdline_matches_config(
 ) -> bool:
     """Check if command line matches our expected configuration."""
     cmdline_str = " ".join(cmdline)
-    print(f"DEBUG: Found process on port {port}: {cmdline_str}")
+    logger.debug(f"Found process on port {port}: {cmdline_str}")
 
     # Check if it's our embedding server
     is_embedding_server = any(
@@ -74,7 +100,7 @@ def _check_cmdline_matches_config(
     )
 
     if not is_embedding_server:
-        print(f"DEBUG: Process on port {port} is not our embedding server")
+        logger.debug(f"Process on port {port} is not our embedding server")
         return False
 
     # Check model name
@@ -84,8 +110,8 @@ def _check_cmdline_matches_config(
     passages_matches = _check_passages_in_cmdline(cmdline, expected_passages_file)
 
     result = model_matches and passages_matches
-    print(
-        f"DEBUG: model_matches: {model_matches}, passages_matches: {passages_matches}, overall: {result}"
+    logger.debug(
+        f"model_matches: {model_matches}, passages_matches: {passages_matches}, overall: {result}"
     )
     return result
 
@@ -132,10 +158,10 @@ def _find_compatible_port_or_next_available(
 
         # Port is in use, check if it's compatible
         if _check_process_matches_config(port, model_name, passages_file):
-            print(f"✅ Found compatible server on port {port}")
+            logger.info(f"Found compatible server on port {port}")
             return port, True
         else:
-            print(f"⚠️  Port {port} has incompatible server, trying next port...")
+            logger.info(f"Port {port} has incompatible server, trying next port...")
 
     raise RuntimeError(
         f"Could not find compatible or available port in range {start_port}-{start_port + max_attempts}"
@@ -156,8 +182,8 @@ class EmbeddingServerManager:
                                        e.g., "leann_backend_diskann.embedding_server"
         """
         self.backend_module_name = backend_module_name
-        self.server_process: Optional[subprocess.Popen] = None
-        self.server_port: Optional[int] = None
+        self.server_process: subprocess.Popen | None = None
+        self.server_port: int | None = None
         self._atexit_registered = False
 
     def start_server(
@@ -167,73 +193,76 @@ class EmbeddingServerManager:
         embedding_mode: str = "sentence-transformers",
         **kwargs,
     ) -> tuple[bool, int]:
-        """
-        Starts the embedding server process.
-
-        Args:
-            port (int): The preferred ZMQ port for the server.
-            model_name (str): The name of the embedding model to use.
-            **kwargs: Additional arguments for the server.
-
-        Returns:
-            tuple[bool, int]: (success, actual_port_used)
-        """
+        """Start the embedding server."""
         passages_file = kwargs.get("passages_file")
-        assert isinstance(passages_file, str), "passages_file must be a string"
 
-        # Check if we have a compatible running server
+        # Check if we have a compatible server already running
         if self._has_compatible_running_server(model_name, passages_file):
-            assert self.server_port is not None, (
-                "a compatible running server should set server_port"
-            )
-            return True, self.server_port
+            logger.info("Found compatible running server!")
+            return True, port
 
-        # Find available port (compatible or free)
-        try:
-            actual_port, is_compatible = _find_compatible_port_or_next_available(
-                port, model_name, passages_file
-            )
-        except RuntimeError as e:
-            print(f"❌ {e}")
-            return False, port
+        # For Colab environment, use a different strategy
+        if _is_colab_environment():
+            logger.info("Detected Colab environment, using alternative startup strategy")
+            return self._start_server_colab(port, model_name, embedding_mode, **kwargs)
+
+        # Find a compatible port or next available
+        actual_port, is_compatible = _find_compatible_port_or_next_available(
+            port, model_name, passages_file
+        )
 
         if is_compatible:
-            print(f"✅ Using existing compatible server on port {actual_port}")
-            self.server_port = actual_port
-            self.server_process = None  # We don't own this process
+            logger.info(f"Found compatible server on port {actual_port}")
             return True, actual_port
 
-        if actual_port != port:
-            print(f"⚠️  Using port {actual_port} instead of {port}")
-
-        # Start new server
+        # Start a new server
         return self._start_new_server(actual_port, model_name, embedding_mode, **kwargs)
 
-    def _has_compatible_running_server(
-        self, model_name: str, passages_file: str
-    ) -> bool:
+    def _start_server_colab(
+        self,
+        port: int,
+        model_name: str,
+        embedding_mode: str = "sentence-transformers",
+        **kwargs,
+    ) -> tuple[bool, int]:
+        """Start server with Colab-specific configuration."""
+        # Try to find an available port
+        try:
+            actual_port = _get_available_port(port)
+        except RuntimeError:
+            logger.error("No available ports found")
+            return False, port
+
+        logger.info(f"Starting server on port {actual_port} for Colab environment")
+
+        # Use a simpler startup strategy for Colab
+        command = self._build_server_command(actual_port, model_name, embedding_mode, **kwargs)
+
+        try:
+            # In Colab, we'll use a more direct approach
+            self._launch_server_process_colab(command, actual_port)
+            return self._wait_for_server_ready_colab(actual_port)
+        except Exception as e:
+            logger.error(f"Failed to start embedding server in Colab: {e}")
+            return False, actual_port
+
+    def _has_compatible_running_server(self, model_name: str, passages_file: str) -> bool:
         """Check if we have a compatible running server."""
-        if not (
-            self.server_process
-            and self.server_process.poll() is None
-            and self.server_port
-        ):
+        if not (self.server_process and self.server_process.poll() is None and self.server_port):
             return False
 
         if _check_process_matches_config(self.server_port, model_name, passages_file):
-            print(
-                f"✅ Existing server process (PID {self.server_process.pid}) is compatible"
-            )
+            logger.info(f"Existing server process (PID {self.server_process.pid}) is compatible")
             return True
 
-        print("⚠️  Existing server process is incompatible. Should start a new server.")
+        logger.info("Existing server process is incompatible. Should start a new server.")
         return False
 
     def _start_new_server(
         self, port: int, model_name: str, embedding_mode: str, **kwargs
     ) -> tuple[bool, int]:
         """Start a new embedding server on the given port."""
-        print(f"INFO: Starting embedding server on port {port}...")
+        logger.info(f"Starting embedding server on port {port}...")
 
         command = self._build_server_command(port, model_name, embedding_mode, **kwargs)
 
@@ -241,7 +270,7 @@ class EmbeddingServerManager:
             self._launch_server_process(command, port)
             return self._wait_for_server_ready(port)
         except Exception as e:
-            print(f"❌ ERROR: Failed to start embedding server: {e}")
+            logger.error(f"Failed to start embedding server: {e}")
             return False, port
 
     def _build_server_command(
@@ -259,29 +288,31 @@ class EmbeddingServerManager:
         ]
 
         if kwargs.get("passages_file"):
-            command.extend(["--passages-file", str(kwargs["passages_file"])])
+            # Convert to absolute path to ensure subprocess can find the file
+            passages_file = Path(kwargs["passages_file"]).resolve()
+            command.extend(["--passages-file", str(passages_file)])
         if embedding_mode != "sentence-transformers":
             command.extend(["--embedding-mode", embedding_mode])
+        if kwargs.get("distance_metric"):
+            command.extend(["--distance-metric", kwargs["distance_metric"]])
 
         return command
 
     def _launch_server_process(self, command: list, port: int) -> None:
         """Launch the server process."""
         project_root = Path(__file__).parent.parent.parent.parent.parent
-        print(f"INFO: Command: {' '.join(command)}")
+        logger.info(f"Command: {' '.join(command)}")
 
+        # Let server output go directly to console
+        # The server will respect LEANN_LOG_LEVEL environment variable
         self.server_process = subprocess.Popen(
             command,
             cwd=project_root,
-            stdout=subprocess.PIPE,
-            stderr=subprocess.STDOUT,
-            text=True,
-            encoding="utf-8",
-            bufsize=1,
-            universal_newlines=True,
+            stdout=None,  # Direct to console
+            stderr=None,  # Direct to console
         )
         self.server_port = port
-        print(f"INFO: Server process started with PID: {self.server_process.pid}")
+        logger.info(f"Server process started with PID: {self.server_process.pid}")
 
         # Register atexit callback only when we actually start a process
         if not self._atexit_registered:
@@ -294,49 +325,19 @@ class EmbeddingServerManager:
         max_wait, wait_interval = 120, 0.5
         for _ in range(int(max_wait / wait_interval)):
             if _check_port(port):
-                print("✅ Embedding server is ready!")
-                threading.Thread(target=self._log_monitor, daemon=True).start()
+                logger.info("Embedding server is ready!")
                 return True, port
 
-            if self.server_process.poll() is not None:
-                print("❌ ERROR: Server terminated during startup.")
-                self._print_recent_output()
+            if self.server_process and self.server_process.poll() is not None:
+                logger.error("Server terminated during startup.")
                 return False, port
 
             time.sleep(wait_interval)
 
-        print(f"❌ ERROR: Server failed to start within {max_wait} seconds.")
+        logger.error(f"Server failed to start within {max_wait} seconds.")
         self.stop_server()
         return False, port
 
-    def _print_recent_output(self):
-        """Print any recent output from the server process."""
-        if not self.server_process or not self.server_process.stdout:
-            return
-        try:
-            if select.select([self.server_process.stdout], [], [], 0)[0]:
-                output = self.server_process.stdout.read()
-                if output:
-                    print(f"[{self.backend_module_name} OUTPUT]: {output}")
-        except Exception as e:
-            print(f"Error reading server output: {e}")
-
-    def _log_monitor(self):
-        """Monitors and prints the server's stdout and stderr."""
-        if not self.server_process:
-            return
-        try:
-            if self.server_process.stdout:
-                while True:
-                    line = self.server_process.stdout.readline()
-                    if not line:
-                        break
-                    print(
-                        f"[{self.backend_module_name} LOG]: {line.strip()}", flush=True
-                    )
-        except Exception as e:
-            print(f"Log monitor error: {e}")
-
     def stop_server(self):
         """Stops the embedding server process if it's running."""
         if not self.server_process:
@@ -347,18 +348,74 @@ class EmbeddingServerManager:
             self.server_process = None
             return
 
-        print(
-            f"INFO: Terminating server process (PID: {self.server_process.pid}) for backend {self.backend_module_name}..."
+        logger.info(
+            f"Terminating server process (PID: {self.server_process.pid}) for backend {self.backend_module_name}..."
         )
         self.server_process.terminate()
 
         try:
-            self.server_process.wait(timeout=5)
-            print(f"INFO: Server process {self.server_process.pid} terminated.")
+            self.server_process.wait(timeout=3)
+            logger.info(f"Server process {self.server_process.pid} terminated.")
         except subprocess.TimeoutExpired:
-            print(
-                f"WARNING: Server process {self.server_process.pid} did not terminate gracefully, killing it."
+            logger.warning(
+                f"Server process {self.server_process.pid} did not terminate gracefully within 3 seconds, killing it."
             )
             self.server_process.kill()
+            try:
+                self.server_process.wait(timeout=2)
+                logger.info(f"Server process {self.server_process.pid} killed successfully.")
+            except subprocess.TimeoutExpired:
+                logger.error(
+                    f"Failed to kill server process {self.server_process.pid} - it may be hung"
+                )
+                # Don't hang indefinitely
+
+        # Clean up process resources to prevent resource tracker warnings
+        try:
+            self.server_process.wait()  # Ensure process is fully cleaned up
+        except Exception:
+            pass
 
         self.server_process = None
+
+    def _launch_server_process_colab(self, command: list, port: int) -> None:
+        """Launch the server process with Colab-specific settings."""
+        logger.info(f"Colab Command: {' '.join(command)}")
+
+        # In Colab, we need to be more careful about process management
+        self.server_process = subprocess.Popen(
+            command,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            text=True,
+        )
+        self.server_port = port
+        logger.info(f"Colab server process started with PID: {self.server_process.pid}")
+
+        # Register atexit callback
+        if not self._atexit_registered:
+            atexit.register(lambda: self.stop_server() if self.server_process else None)
+            self._atexit_registered = True
+
+    def _wait_for_server_ready_colab(self, port: int) -> tuple[bool, int]:
+        """Wait for the server to be ready with Colab-specific timeout."""
+        max_wait, wait_interval = 30, 0.5  # Shorter timeout for Colab
+
+        for _ in range(int(max_wait / wait_interval)):
+            if _check_port(port):
+                logger.info("Colab embedding server is ready!")
+                return True, port
+
+            if self.server_process and self.server_process.poll() is not None:
+                # Check for error output
+                stdout, stderr = self.server_process.communicate()
+                logger.error("Colab server terminated during startup.")
+                logger.error(f"stdout: {stdout}")
+                logger.error(f"stderr: {stderr}")
+                return False, port
+
+            time.sleep(wait_interval)
+
+        logger.error(f"Colab server failed to start within {max_wait} seconds.")
+        self.stop_server()
+        return False, port

packages/leann-core/src/leann/interface.py

@@ -1,15 +1,14 @@
 from abc import ABC, abstractmethod
+from typing import Any, Literal
+
 import numpy as np
-from typing import Dict, Any, List, Literal
 
 
 class LeannBackendBuilderInterface(ABC):
     """Backend interface for building indexes"""
 
     @abstractmethod
-    def build(
-        self, data: np.ndarray, ids: List[str], index_path: str, **kwargs
-    ) -> None:
+    def build(self, data: np.ndarray, ids: list[str], index_path: str, **kwargs) -> None:
         """Build index
 
         Args:
@@ -34,6 +33,11 @@ class LeannBackendSearcherInterface(ABC):
         """
         pass
 
+    @abstractmethod
+    def _ensure_server_running(self, passages_source_file: str, port: int | None, **kwargs) -> int:
+        """Ensure server is running"""
+        pass
+
     @abstractmethod
     def search(
         self,
@@ -44,9 +48,9 @@ class LeannBackendSearcherInterface(ABC):
         prune_ratio: float = 0.0,
         recompute_embeddings: bool = False,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: int = 5557,
+        zmq_port: int | None = None,
         **kwargs,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         """Search for nearest neighbors
 
         Args:
@@ -57,7 +61,7 @@ class LeannBackendSearcherInterface(ABC):
             prune_ratio: Ratio of neighbors to prune via approximate distance (0.0-1.0)
             recompute_embeddings: Whether to fetch fresh embeddings from server vs use stored PQ codes
             pruning_strategy: PQ candidate selection strategy - "global" (default), "local", or "proportional"
-            zmq_port: ZMQ port for embedding server communication
+            zmq_port: ZMQ port for embedding server communication. Must be provided if recompute_embeddings is True.
             **kwargs: Backend-specific parameters
 
         Returns:
@@ -67,7 +71,10 @@ class LeannBackendSearcherInterface(ABC):
 
     @abstractmethod
     def compute_query_embedding(
-        self, query: str, zmq_port: int = 5557, use_server_if_available: bool = True
+        self,
+        query: str,
+        use_server_if_available: bool = True,
+        zmq_port: int | None = None,
     ) -> np.ndarray:
         """Compute embedding for a query string
 

packages/leann-core/src/leann/registry.py

@@ -1,13 +1,13 @@
 # packages/leann-core/src/leann/registry.py
 
-from typing import Dict, TYPE_CHECKING
 import importlib
 import importlib.metadata
+from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     from leann.interface import LeannBackendFactoryInterface
 
-BACKEND_REGISTRY: Dict[str, "LeannBackendFactoryInterface"] = {}
+BACKEND_REGISTRY: dict[str, "LeannBackendFactoryInterface"] = {}
 
 
 def register_backend(name: str):
@@ -31,13 +31,11 @@ def autodiscover_backends():
             backend_module_name = dist_name.replace("-", "_")
             discovered_backends.append(backend_module_name)
 
-    for backend_module_name in sorted(
-        discovered_backends
-    ):  # sort for deterministic loading
+    for backend_module_name in sorted(discovered_backends):  # sort for deterministic loading
         try:
             importlib.import_module(backend_module_name)
             # Registration message is printed by the decorator
-        except ImportError as e:
+        except ImportError:
             # print(f"WARN: Could not import backend module '{backend_module_name}': {e}")
             pass
     # print("INFO: Backend auto-discovery finished.")

packages/leann-core/src/leann/searcher_base.py

@@ -1,8 +1,7 @@
 import json
-import pickle
 from abc import ABC, abstractmethod
 from pathlib import Path
-from typing import Dict, Any, Literal
+from typing import Any, Literal
 
 import numpy as np
 
@@ -39,54 +38,56 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
 
         self.embedding_model = self.meta.get("embedding_model")
         if not self.embedding_model:
-            print(
-                "WARNING: embedding_model not found in meta.json. Recompute will fail."
-            )
+            print("WARNING: embedding_model not found in meta.json. Recompute will fail.")
+
+        self.embedding_mode = self.meta.get("embedding_mode", "sentence-transformers")
 
         self.embedding_server_manager = EmbeddingServerManager(
-            backend_module_name=backend_module_name
+            backend_module_name=backend_module_name,
         )
 
-    def _load_meta(self) -> Dict[str, Any]:
+    def _load_meta(self) -> dict[str, Any]:
         """Loads the metadata file associated with the index."""
         # This is the corrected logic for finding the meta file.
         meta_path = self.index_dir / f"{self.index_path.name}.meta.json"
         if not meta_path.exists():
             raise FileNotFoundError(f"Leann metadata file not found at {meta_path}")
-        with open(meta_path, "r", encoding="utf-8") as f:
+        with open(meta_path, encoding="utf-8") as f:
             return json.load(f)
 
-    def _ensure_server_running(
-        self, passages_source_file: str, port: int, **kwargs
-    ) -> int:
+    def _ensure_server_running(self, passages_source_file: str, port: int, **kwargs) -> int:
         """
         Ensures the embedding server is running if recompute is needed.
         This is a helper for subclasses.
         """
         if not self.embedding_model:
-            raise ValueError(
-                "Cannot use recompute mode without 'embedding_model' in meta.json."
-            )
+            raise ValueError("Cannot use recompute mode without 'embedding_model' in meta.json.")
 
-        embedding_mode = self.meta.get("embedding_mode", "sentence-transformers")
+        # Get distance_metric from meta if not provided in kwargs
+        distance_metric = (
+            kwargs.get("distance_metric")
+            or self.meta.get("backend_kwargs", {}).get("distance_metric")
+            or "mips"
+        )
 
         server_started, actual_port = self.embedding_server_manager.start_server(
             port=port,
             model_name=self.embedding_model,
+            embedding_mode=self.embedding_mode,
             passages_file=passages_source_file,
-            distance_metric=kwargs.get("distance_metric"),
-            embedding_mode=embedding_mode,
+            distance_metric=distance_metric,
             enable_warmup=kwargs.get("enable_warmup", False),
         )
         if not server_started:
-            raise RuntimeError(
-                f"Failed to start embedding server on port {actual_port}"
-            )
+            raise RuntimeError(f"Failed to start embedding server on port {actual_port}")
 
         return actual_port
 
     def compute_query_embedding(
-        self, query: str, zmq_port: int = 5557, use_server_if_available: bool = True
+        self,
+        query: str,
+        use_server_if_available: bool = True,
+        zmq_port: int = 5557,
     ) -> np.ndarray:
         """
         Compute embedding for a query string.
@@ -102,12 +103,15 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
         # Try to use embedding server if available and requested
         if use_server_if_available:
             try:
+                # TODO: Maybe we can directly use this port here?
+                # For this internal method, it's ok to assume that the server is running
+                # on that port?
+
                 # Ensure we have a server with passages_file for compatibility
-                passages_source_file = (
-                    self.index_dir / f"{self.index_path.name}.meta.json"
-                )
+                passages_source_file = self.index_dir / f"{self.index_path.name}.meta.json"
+                # Convert to absolute path to ensure server can find it
                 zmq_port = self._ensure_server_running(
-                    str(passages_source_file), zmq_port
+                    str(passages_source_file.resolve()), zmq_port
                 )
 
                 return self._compute_embedding_via_server([query], zmq_port)[
@@ -118,15 +122,15 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
                 print("⏭️ Falling back to direct model loading...")
 
         # Fallback to direct computation
-        from .api import compute_embeddings
+        from .embedding_compute import compute_embeddings
 
         embedding_mode = self.meta.get("embedding_mode", "sentence-transformers")
         return compute_embeddings([query], self.embedding_model, embedding_mode)
 
     def _compute_embedding_via_server(self, chunks: list, zmq_port: int) -> np.ndarray:
         """Compute embeddings using the ZMQ embedding server."""
-        import zmq
         import msgpack
+        import zmq
 
         try:
             context = zmq.Context()
@@ -165,9 +169,9 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
         prune_ratio: float = 0.0,
         recompute_embeddings: bool = False,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: int = 5557,
+        zmq_port: int | None = None,
         **kwargs,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         """
         Search for the top_k nearest neighbors of the query vector.
 
@@ -179,7 +183,7 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
             prune_ratio: Ratio of neighbors to prune via approximate distance (0.0-1.0)
             recompute_embeddings: Whether to fetch fresh embeddings from server vs use stored PQ codes
             pruning_strategy: PQ candidate selection strategy - "global" (default), "local", or "proportional"
-            zmq_port: ZMQ port for embedding server communication
+            zmq_port: ZMQ port for embedding server communication. Must be provided if recompute_embeddings is True.
             **kwargs: Backend-specific parameters (e.g., batch_size, dedup_node_dis, etc.)
 
         Returns:

packages/leann/README.md

@@ -0,0 +1,36 @@
+# LEANN - The smallest vector index in the world
+
+LEANN is a revolutionary vector database that democratizes personal AI. Transform your laptop into a powerful RAG system that can index and search through millions of documents while using **97% less storage** than traditional solutions **without accuracy loss**.
+
+## Installation
+
+```bash
+# Default installation (includes both HNSW and DiskANN backends)
+uv pip install leann
+```
+
+## Quick Start
+
+```python
+from leann import LeannBuilder, LeannSearcher, LeannChat
+from pathlib import Path
+INDEX_PATH = str(Path("./").resolve() / "demo.leann")
+
+# Build an index (choose backend: "hnsw" or "diskann")
+builder = LeannBuilder(backend_name="hnsw")  # or "diskann" for large-scale deployments
+builder.add_text("LEANN saves 97% storage compared to traditional vector databases.")
+builder.add_text("Tung Tung Tung Sahur called—they need their banana‑crocodile hybrid back")
+builder.build_index(INDEX_PATH)
+
+# Search
+searcher = LeannSearcher(INDEX_PATH)
+results = searcher.search("fantastical AI-generated creatures", top_k=1)
+
+# Chat with your data
+chat = LeannChat(INDEX_PATH, llm_config={"type": "hf", "model": "Qwen/Qwen3-0.6B"})
+response = chat.ask("How much storage does LEANN save?", top_k=1)
+```
+
+## License
+
+MIT License

packages/leann/__init__.py

@@ -0,0 +1,12 @@
+"""
+LEANN - Low-storage Embedding Approximation for Neural Networks
+
+A revolutionary vector database that democratizes personal AI.
+"""
+
+__version__ = "0.1.0"
+
+# Re-export main API from leann-core
+from leann_core import LeannBuilder, LeannChat, LeannSearcher
+
+__all__ = ["LeannBuilder", "LeannChat", "LeannSearcher"]

packages/leann/pyproject.toml

@@ -0,0 +1,39 @@
+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "leann"
+version = "0.1.16"
+description = "LEANN - The smallest vector index in the world. RAG Everything with LEANN!"
+readme = "README.md"
+requires-python = ">=3.9"
+license = { text = "MIT" }
+authors = [
+    { name = "LEANN Team" }
+]
+keywords = ["vector-database", "rag", "embeddings", "search", "ai"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+]
+
+# Default installation: core + hnsw + diskann
+dependencies = [
+    "leann-core>=0.1.0",
+    "leann-backend-hnsw>=0.1.0",
+    "leann-backend-diskann>=0.1.0",
+]
+
+[project.optional-dependencies]
+# All backends now included by default
+
+[project.urls]
+Repository = "https://github.com/yichuan-w/LEANN"
+Issues = "https://github.com/yichuan-w/LEANN/issues"

packages/wechat-exporter/main.py

@@ -1,22 +1,23 @@
 import json
-import typer
-from pathlib import Path
-import requests
-from tqdm import tqdm
-import xml.etree.ElementTree as ET
-from typing_extensions import Annotated
 import sqlite3
+import xml.etree.ElementTree as ElementTree
+from pathlib import Path
+from typing import Annotated
+
+import requests
+import typer
+from tqdm import tqdm
 
 app = typer.Typer()
 
+
 def get_safe_path(s: str) -> str:
     """
     Remove invalid characters to sanitize a path.
     :param s: str to sanitize
     :returns: sanitized str
     """
-    ban_chars = "\\  /  :  *  ?  \"  '  <  >  |  $  \r  \n".replace(
-        ' ', '')
+    ban_chars = "\\  /  :  *  ?  \"  '  <  >  |  $  \r  \n".replace(" ", "")
     for i in ban_chars:
         s = s.replace(i, "")
     return s
@@ -25,36 +26,40 @@ def get_safe_path(s: str) -> str:
 def process_history(history: str):
     if history.startswith("<?xml") or history.startswith("<msg>"):
         try:
-            root = ET.fromstring(history)
-            title = root.find('.//title').text if root.find('.//title') is not None else None
-            quoted = root.find('.//refermsg/content').text if root.find('.//refermsg/content') is not None else None
+            root = ElementTree.fromstring(history)
+            title = root.find(".//title").text if root.find(".//title") is not None else None
+            quoted = (
+                root.find(".//refermsg/content").text
+                if root.find(".//refermsg/content") is not None
+                else None
+            )
             if title and quoted:
-                return {
-                    "title": title,
-                    "quoted": process_history(quoted)
-                }
+                return {"title": title, "quoted": process_history(quoted)}
             if title:
                 return title
         except Exception:
             return history
     return history
 
+
 def get_message(history: dict | str):
     if isinstance(history, dict):
-        if 'title' in history:
-            return history['title']
+        if "title" in history:
+            return history["title"]
     else:
         return history
 
+
 def export_chathistory(user_id: str):
-    res = requests.get("http://localhost:48065/wechat/chatlog", params={
-        "userId": user_id,
-        "count": 100000
-    }).json()
-    for i in range(len(res['chatLogs'])):
-        res['chatLogs'][i]['content'] = process_history(res['chatLogs'][i]['content'])
-        res['chatLogs'][i]['message'] = get_message(res['chatLogs'][i]['content'])
-    return res['chatLogs']
+    res = requests.get(
+        "http://localhost:48065/wechat/chatlog",
+        params={"userId": user_id, "count": 100000},
+    ).json()
+    for i in range(len(res["chatLogs"])):
+        res["chatLogs"][i]["content"] = process_history(res["chatLogs"][i]["content"])
+        res["chatLogs"][i]["message"] = get_message(res["chatLogs"][i]["content"])
+    return res["chatLogs"]
+
 
 @app.command()
 def export_all(dest: Annotated[Path, typer.Argument(help="Destination path to export to.")]):
@@ -64,7 +69,7 @@ def export_all(dest: Annotated[Path, typer.Argument(help="Destination path to ex
     if not dest.is_dir():
         if not dest.exists():
             inp = typer.prompt("Destination path does not exist, create it? (y/n)")
-            if inp.lower() == 'y':
+            if inp.lower() == "y":
                 dest.mkdir(parents=True)
             else:
                 typer.echo("Aborted.", err=True)
@@ -77,12 +82,12 @@ def export_all(dest: Annotated[Path, typer.Argument(help="Destination path to ex
     exported_count = 0
     for user in tqdm(all_users):
         try:
-            usr_chatlog = export_chathistory(user['arg'])
-            
+            usr_chatlog = export_chathistory(user["arg"])
+
             # Only write file if there are messages
             if len(usr_chatlog) > 0:
-                out_path = dest/get_safe_path((user['title'] or "")+"-"+user['arg']+'.json')
-                with open(out_path, 'w', encoding='utf-8') as f:
+                out_path = dest / get_safe_path((user["title"] or "") + "-" + user["arg"] + ".json")
+                with open(out_path, "w", encoding="utf-8") as f:
                     json.dump(usr_chatlog, f, ensure_ascii=False, indent=2)
                 exported_count += 1
         except Exception as e:
@@ -91,23 +96,43 @@ def export_all(dest: Annotated[Path, typer.Argument(help="Destination path to ex
 
     print(f"Exported {exported_count} users' chat history to {dest} in json.")
 
+
 @app.command()
-def export_sqlite(dest: Annotated[Path, typer.Argument(help="Destination path to export to.")] = Path("chatlog.db")):
+def export_sqlite(
+    dest: Annotated[Path, typer.Argument(help="Destination path to export to.")] = Path(
+        "chatlog.db"
+    ),
+):
     """
     Export all users' chat history to a sqlite database.
     """
     connection = sqlite3.connect(dest)
     cursor = connection.cursor()
-    cursor.execute("CREATE TABLE IF NOT EXISTS chatlog (id INTEGER PRIMARY KEY AUTOINCREMENT, with_id TEXT, from_user TEXT, to_user TEXT, message TEXT, timest DATETIME, auxiliary TEXT)")
+    cursor.execute(
+        "CREATE TABLE IF NOT EXISTS chatlog (id INTEGER PRIMARY KEY AUTOINCREMENT, with_id TEXT, from_user TEXT, to_user TEXT, message TEXT, timest DATETIME, auxiliary TEXT)"
+    )
     cursor.execute("CREATE INDEX IF NOT EXISTS chatlog_with_id_index ON chatlog (with_id)")
     cursor.execute("CREATE TABLE iF NOT EXISTS users (id TEXT PRIMARY KEY, name TEXT)")
 
     all_users = requests.get("http://localhost:48065/wechat/allcontacts").json()
     for user in tqdm(all_users):
-        cursor.execute("INSERT OR IGNORE INTO users (id, name) VALUES (?, ?)", (user['arg'], user['title']))
-        usr_chatlog = export_chathistory(user['arg'])
+        cursor.execute(
+            "INSERT OR IGNORE INTO users (id, name) VALUES (?, ?)",
+            (user["arg"], user["title"]),
+        )
+        usr_chatlog = export_chathistory(user["arg"])
         for msg in usr_chatlog:
-            cursor.execute("INSERT INTO chatlog (with_id, from_user, to_user, message, timest, auxiliary) VALUES (?, ?, ?, ?, ?, ?)", (user['arg'], msg['fromUser'], msg['toUser'], msg['message'], msg['createTime'], str(msg['content'])))
+            cursor.execute(
+                "INSERT INTO chatlog (with_id, from_user, to_user, message, timest, auxiliary) VALUES (?, ?, ?, ?, ?, ?)",
+                (
+                    user["arg"],
+                    msg["fromUser"],
+                    msg["toUser"],
+                    msg["message"],
+                    msg["createTime"],
+                    str(msg["content"]),
+                ),
+            )
     connection.commit()
 
 

pyproject.toml

@@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta"
 [project]
 name = "leann-workspace"
 version = "0.1.0"
-requires-python = ">=3.10"
+requires-python = ">=3.9"
 
 dependencies = [
     "leann-core",
@@ -25,16 +25,23 @@ dependencies = [
     "requests>=2.25.0",
     "sentence-transformers>=2.2.0",
     "openai>=1.0.0",
+    # PDF parsing dependencies - essential for document processing
     "PyPDF2>=3.0.0",
+    "pdfplumber>=0.11.0",
+    "pymupdf>=1.26.0",
+    "pypdfium2>=4.30.0",
+    # LlamaIndex core and readers - updated versions
     "llama-index>=0.12.44",
-    "llama-index-readers-docling",
-    "llama-index-node-parser-docling",
-    "ipykernel==6.29.5",
-    "msgpack>=1.1.1",
+    "llama-index-readers-file>=0.4.0",  # Essential for PDF parsing
+    # "llama-index-readers-docling",  # Requires Python >= 3.10
+    # "llama-index-node-parser-docling",  # Requires Python >= 3.10
     "llama-index-vector-stores-faiss>=0.4.0",
     "llama-index-embeddings-huggingface>=0.5.5",
-    "mlx>=0.26.3",
-    "mlx-lm>=0.26.0",
+    # Other dependencies
+    "ipykernel==6.29.5",
+    "msgpack>=1.1.1",
+    "mlx>=0.26.3; sys_platform == 'darwin'",
+    "mlx-lm>=0.26.0; sys_platform == 'darwin'",
     "psutil>=5.8.0",
 ]
 
@@ -42,16 +49,35 @@ dependencies = [
 dev = [
     "pytest>=7.0",
     "pytest-cov>=4.0",
+    "pytest-xdist>=3.0",  # For parallel test execution
     "black>=23.0",
     "ruff>=0.1.0",
     "matplotlib",
     "huggingface-hub>=0.20.0",
+    "pre-commit>=3.5.0",
+]
+
+test = [
+    "pytest>=7.0",
+    "pytest-timeout>=2.0",
+    "llama-index-core>=0.12.0",
+    "llama-index-readers-file>=0.4.0",
+    "python-dotenv>=1.0.0",
+    "sentence-transformers>=2.2.0",
 ]
 
 diskann = [
     "leann-backend-diskann",
 ]
 
+# Add a new optional dependency group for document processing
+documents = [
+    "beautifulsoup4>=4.13.0",  # For HTML parsing
+    "python-docx>=0.8.11",     # For Word documents
+    "openpyxl>=3.1.0",         # For Excel files
+    "pandas>=2.2.0",           # For data processing
+]
+
 [tool.setuptools]
 py-modules = []
 
@@ -60,3 +86,71 @@ py-modules = []
 leann-core = { path = "packages/leann-core", editable = true }
 leann-backend-diskann = { path = "packages/leann-backend-diskann", editable = true }
 leann-backend-hnsw = { path = "packages/leann-backend-hnsw", editable = true }
+
+[tool.ruff]
+target-version = "py310"
+line-length = 100
+extend-exclude = [
+    "third_party",
+    "*.egg-info",
+    "__pycache__",
+    ".git",
+    ".venv",
+]
+
+[tool.ruff.lint]
+select = [
+    "E",      # pycodestyle errors
+    "W",      # pycodestyle warnings
+    "F",      # pyflakes
+    "I",      # isort
+    "B",      # flake8-bugbear
+    "C4",     # flake8-comprehensions
+    "UP",     # pyupgrade
+    "N",      # pep8-naming
+    "RUF",    # ruff-specific rules
+]
+ignore = [
+    "E501",   # line too long (handled by formatter)
+    "B008",   # do not perform function calls in argument defaults
+    "B904",   # raise without from
+    "N812",   # lowercase imported as non-lowercase
+    "N806",   # variable in function should be lowercase
+    "RUF012", # mutable class attributes should be annotated with typing.ClassVar
+]
+
+[tool.ruff.lint.per-file-ignores]
+"test/**/*.py" = ["E402"]      # module level import not at top of file (common in tests)
+"examples/**/*.py" = ["E402"]  # module level import not at top of file (common in examples)
+
+[tool.ruff.format]
+quote-style = "double"
+indent-style = "space"
+skip-magic-trailing-comma = false
+line-ending = "auto"
+
+[dependency-groups]
+dev = [
+    "ruff>=0.12.4",
+]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+python_files = ["test_*.py"]
+python_classes = ["Test*"]
+python_functions = ["test_*"]
+markers = [
+    "slow: marks tests as slow (deselect with '-m \"not slow\"')",
+    "openai: marks tests that require OpenAI API key",
+]
+timeout = 600
+addopts = [
+    "-v",
+    "--tb=short",
+    "--strict-markers",
+    "--disable-warnings",
+]
+env = [
+    "HF_HUB_DISABLE_SYMLINKS=1",
+    "TOKENIZERS_PARALLELISM=false",
+]

research/micro/analyze_HNSW.py

@@ -1,12 +0,0 @@
-import faiss
-hnsw_index = faiss.read_index("/opt/dlami/nvme/scaling_out/indices/rpj_wiki/facebook/contriever-msmarco/hnsw/hnsw_IP_M30_efC128.index", faiss.IO_FLAG_ONDISK_SAME_DIR)
-
-# print total number of nodes
-print(hnsw_index.ntotal)
-
-# print  stats of the graph
-print(hnsw_index.hnsw.print_neighbor_stats(0))
-
-
-# save_degree_distribution
-hnsw_index.hnsw.save_degree_distribution(0, "degree_distribution_HNSW_M30.txt")

research/micro/analyze_NSG.py

@@ -1,11 +0,0 @@
-import faiss
-nsg_index = faiss.read_index("/opt/dlami/nvme/scaling_out/indices/rpj_wiki/facebook/contriever-msmarco/nsg_R16.index", faiss.IO_FLAG_ONDISK_SAME_DIR)
-
-# print total number of nodes
-print(nsg_index.ntotal)
-
-# print stats of the graph
-print(nsg_index.nsg.print_neighbor_stats(0))
-
-# save degree distribution
-nsg_index.nsg.save_degree_distribution("degree_distribution_NSG_R60.txt")

research/micro/bnbtest.py

@@ -1,63 +0,0 @@
-import torch
-import torch.nn as nn
-import time
-
-# import bitsandbytes as bnb
-from bitsandbytes.nn import Linear8bitLt
-
-# set default to half
-import torch
-torch.set_default_dtype(torch.float16)
-
-M = 2048
-N = 2048
-
-bsz =  2048
-import torch_int
-from torch_int.nn.linear import W8A8BFP32OFP32Linear, W8A8B8O8Linear, W8A8B8O8LinearReLU
-
-fp16_model = nn.Sequential(
-    nn.Linear(M, N),
-    # nn.Linear(2048, 2048)
-)
-
-int8_model = nn.Sequential(
-    Linear8bitLt(M, N, has_fp16_weights=False),
-    # Linear8bitLt(2048, 2048, has_fp16_weights=False)
-)
-
-int8_model.load_state_dict(fp16_model.state_dict())
-int8_model = int8_model.to(0) # Quantization happens here
-fp16_model = fp16_model.to(0) # Move fp16 model to GPU as well
-
-# Create random input tensor
-input_tensor = torch.randn(bsz, M, device=0)  # Batch of 1000 vectors
-
-# Speed test function
-def speed_test(model, input_tensor, name, num_iterations=100):
-    # Warmup
-    for _ in range(10):
-        _ = model(input_tensor)
-    
-    # Actual timing
-    torch.cuda.synchronize()
-    start_time = time.time()
-    
-    for _ in range(num_iterations):
-        _ = model(input_tensor)
-    
-    torch.cuda.synchronize()
-    end_time = time.time()
-    
-    avg_time = (end_time - start_time) / num_iterations
-    print(f"{name} model: {avg_time:.6f} seconds per iteration")
-    return avg_time
-
-# Run speed tests
-with torch.no_grad():  # Disable gradient calculation for inference
-    fp16_time = speed_test(fp16_model, input_tensor, "FP16")
-    int8_time = speed_test(int8_model, input_tensor, "INT8")
-    
-    # Calculate speedup
-    speedup = fp16_time / int8_time
-    print(f"INT8 is {speedup:.2f}x faster than FP16")

research/micro/data/transformer-batching-microbenchmarks.csv

@@ -1,89 +0,0 @@
-n,d,seqlen,bs,latency,h,flop,io,intensity,throughput,series
-3,256,256,2048,0.009623501679245285,768,618475290624,167.48502132816208,3692720015.912285,64267177503366.266,dense
-3,256,256,1024,0.004853848615384615,768,309237645312,166.15392854317415,1861151572.059558,63709783682138.234,dense
-3,256,256,512,0.0024687246971962615,768,154618822656,163.57953256539062,945221081.3366361,62631051097597.516,dense
-3,256,256,256,0.0012845360838052097,768,77309411328,157.64931990085577,490388486.1451936,60184694149645.54,dense
-3,256,256,128,0.0006901147179878049,768,38654705664,147.57393422494675,261934506.70684624,56012000116019.945,dense
-3,256,256,64,0.0003363830693015702,768,19327352832,153.1328437752606,126212981.84970059,57456378146882.51,dense
-3,256,256,32,0.00018671159748991485,768,9663676416,141.10249365427362,68486928.65540518,51757237075334.75,dense
-3,256,256,16,0.00012353640857142858,768,4831838208,111.40488993609125,43371868.24359184,39112665358133.98,dense
-3,256,256,8,9.774760007849294e-05,768,2415919104,76.43260800265766,31608487.09906635,24715891766754.14,dense
-3,256,256,4,6.672271167474822e-05,768,1207959552,64.82614227498455,18633833.660438772,18104173551704.773,dense
-3,256,256,2,4.9758770289855074e-05,768,603979776,55.317122669351576,10918495.880745342,12138157202874.861,dense
-3,256,1,2048,9.785507940251571e-05,768,2415919104,76.34865809334705,31643242.518371396,24688745017132.86,dense
-3,256,1,1024,6.692813470149253e-05,768,1207959552,64.62717090938949,18691202.70936228,18048606275785.867,dense
-3,256,1,512,4.9680950036205655e-05,768,603979776,55.40377142534654,10901419.893658841,12157170415618.898,dense
-3,256,1,256,4.2781118741058655e-05,768,301989888,45.95672244805227,6571179.83862661,7058952568020.829,dense
-3,256,1,128,5.0662328255350016e-05,768,150994944,31.046026784880404,4863583.512513602,2980418571348.519,dense
-3,256,1,64,4.475009253945481e-05,768,75497472,30.75426042497223,2454862.219307235,1687090857598.4766,dense
-3,256,1,32,4.51682671454219e-05,768,37748736,28.29313765537115,1334201.1218340008,835735758435.5786,dense
-3,256,1,16,5.03585186661834e-05,768,18874368,24.401035466223117,773506.846712577,374799904761.1871,dense
-3,256,1,8,5.023459565217391e-05,768,9437184,23.972005435021096,393675.19858030166,187862246674.45105,dense
-3,256,1,4,5.053219391083726e-05,768,4718592,23.58765586356967,200044.97383259286,93377936614.54384,dense
-3,256,1,2,4.4607398995335484e-05,768,2359296,26.58285456464288,88752.54515134107,52890239133.797226,dense
-12,256,256,2048,0.14480779847058822,3072,9895604649984,44.620009282941716,221775046868.20184,68336130750540.26,dense
-12,256,256,1024,0.07254347629166667,3072,4947802324992,44.664248332585096,110777691547.58836,68204648824643.82,dense
-12,256,256,512,0.036310761444444443,3072,2473901162496,44.876147984203506,55127306456.13385,68131349056975.164,dense
-12,256,256,256,0.01821551906896552,3072,1236950581248,45.24607467289738,27338295977.947884,67906414116709.98,dense
-12,256,256,128,0.009229417903030302,3072,618475290624,45.67217092440895,13541622351.335684,67011299859001.46,dense
-12,256,256,64,0.004754550595394737,3072,309237645312,46.31372736116993,6677019167.566916,65040352207320.695,dense
-12,256,256,32,0.002405752659340659,3072,154618822656,49.68826015254682,3111777755.5766335,64270456921525.82,dense
-12,256,256,16,0.0012287219045005488,3072,77309411328,56.323579604557374,1372594069.3184311,62918558743709.18,dense
-12,256,256,8,0.0006206816149425287,3072,38654705664,70.95456179103653,544781120.315271,62277832520589.78,dense
-12,256,256,4,0.0003875502697142857,3072,19327352832,81.16954743236613,238110885.71245712,49870569942445.75,dense
-12,256,256,2,0.00027502018627941914,3072,9663676416,91.50537035282076,105607751.53129694,35138062215483.168,dense
-12,256,1,2048,0.0006202853873290136,3072,38654705664,70.99988634205897,544433345.6784943,62317614526515.766,dense
-12,256,1,1024,0.00038721467732724153,3072,19327352832,81.2398957010995,237904697.74985722,49913791918755.53,dense
-12,256,1,512,0.000274364799,3072,9663676416,91.72395326121995,105356082.81599998,35221998052308.45,dense
-12,256,1,256,0.00012488918589482266,3072,4831838208,176.31707535146046,27404255.647778228,38689003962834.75,dense
-12,256,1,128,8.976711102514506e-05,3072,2415919104,227.78088507574267,10606329.425740216,26913187652026.21,dense
-12,256,1,64,8.715176287471176e-05,3072,1207959552,225.59268282689945,5354604.31102229,13860414432884.701,dense
-12,256,1,32,8.523013435114503e-05,3072,603979776,226.06539514085782,2671703.8033338524,7086458100741.991,dense
-12,256,1,16,7.901561645904116e-05,3072,301989888,241.35704882952732,1251216.3595988373,3821901309300.556,dense
-12,256,1,8,7.827949114210329e-05,3072,150994944,242.37091635608994,622991.1833900034,1928920867994.581,dense
-12,256,1,4,7.779445951035782e-05,3072,75497472,243.25022783249054,310369.58391664835,970473636235.5986,dense
-12,256,1,2,7.758845406626506e-05,3072,37748736,243.57933441822672,154975.11761480253,486525172518.07056,dense
-3,256,256,2048,0.00507974918466899,768,206158430208,475.59810852303485,433471930.42508715,40584371927298.98,qk_init
-3,256,256,1024,0.0025616677649325623,768,103079215104,471.5519977009198,218595649.27424532,40239103803811.82,qk_init
-3,256,256,512,0.0013029336670480549,768,51539607552,463.55374128015677,111183672.92143403,39556585922573.38,qk_init
-3,256,256,256,0.0006738189029345373,768,25769803776,448.1766342333362,57499213.050413854,38244406121244.69,qk_init
-3,256,256,128,0.000358254672959467,768,12884901888,421.47375986100144,30571065.425874516,35965760841472.125,qk_init
-3,256,256,64,0.0002007051105022831,768,6442450944,376.1611839930762,17126836.096194826,32099087700742.5,qk_init
-3,256,256,32,0.00012189697230142565,768,3221225472,309.6773881032524,10401874.969721656,26425803784810.87,qk_init
-3,256,256,16,8.453561698040722e-05,768,1610612736,223.2711923587723,7213705.982328083,19052475081281.902,qk_init
-3,256,256,8,6.407660705009276e-05,768,805306368,147.2797083750448,5467870.468274581,12567868448003.822,qk_init
-3,256,256,4,5.036328747284576e-05,768,402653184,93.69110391262903,4297667.197682838,7994974200544.344,qk_init
-3,256,256,2,4.5488761135057476e-05,768,201326592,51.865470527877875,3881707.616858238,4425853485045.578,qk_init
-12,256,256,2048,0.020202365999999996,3072,824633720832,478.3437947812648,1723935231.9999998,40818670488001.266,qk_init
-12,256,256,1024,0.010124155888157895,3072,412316860416,477.2583770318811,863927969.1228071,40726048173387.19,qk_init
-12,256,256,512,0.005085633937062937,3072,206158430208,475.04777848703077,433974095.9627039,40537410430893.29,qk_init
-12,256,256,256,0.0025654916853281853,3072,103079215104,470.84913933193053,218921957.14800516,40179126556324.74,qk_init
-12,256,256,128,0.0013045765704467354,3072,51539607552,462.9699702434292,111323867.34478809,39506770794105.96,qk_init
-12,256,256,64,0.0006742801519939804,3072,25769803776,447.87005387442576,57538572.970153,38218244597284.33,qk_init
-12,256,256,32,0.00035831976790671853,3072,12884901888,421.3971919051604,30576620.194706645,35959227042573.69,qk_init
-12,256,256,16,0.0002005369068918302,3072,6442450944,376.4766953382971,17112482.721436176,32126011335534.68,qk_init
-12,256,256,8,0.00012179187250509165,3072,3221225472,309.94462293386505,10392906.453767821,26448607823689.82,qk_init
-12,256,256,4,8.452507263643351e-05,3072,1610612736,223.2990450204527,7212806.198308992,19054851841745.297,qk_init
-12,256,256,2,6.412381767545489e-05,3072,805306368,147.17127491946468,5471899.108305484,12558615459794.32,qk_init
-3,256,256,2048,0.0016183739398395718,768,805306368,811597824.0,0.9922480620155039,1265467.7325087283,qk_ar
-3,256,256,1024,0.0008322699728813558,768,402653184,405798912.0,0.9922480620155039,1230369.9921491416,qk_ar
-3,256,256,512,0.00043886859397590365,768,201326592,202899456.0,0.9922480620155039,1166636.2255762408,qk_ar
-3,256,256,256,0.00024185948322147648,768,100663296,101449728.0,0.9922480620155039,1058465.8355760013,qk_ar
-3,256,256,128,0.00014308985100166944,768,50331648,50724864.0,0.9922480620155039,894542.82818777,qk_ar
-3,256,256,64,9.382939365815932e-05,768,25165824,25362432.0,0.9922480620155039,682089.028872613,qk_ar
-3,256,256,32,6.856070612244899e-05,768,12582912,12681216.0,0.9922480620155039,466739.6503012703,qk_ar
-3,256,256,16,5.452260553129549e-05,768,6291456,6340608.0,0.9922480620155039,293456.26174846216,qk_ar
-3,256,256,8,4.608557533261417e-05,768,3145728,3170304.0,0.9922480620155039,173590.1080166944,qk_ar
-3,256,256,4,4.386146957766642e-05,768,1572864,1585152.0,0.9922480620155039,91196.21477609445,qk_ar
-3,256,256,2,4.330941094420601e-05,768,786432,792576.0,0.9922480620155039,46179.33969539622,qk_ar
-12,256,256,2048,0.006347041645299144,3072,3221225472,3246391296.0,0.9922480620155039,322670.011392918,qk_ar
-12,256,256,1024,0.0031943104467592586,3072,1610612736,1623195648.0,0.9922480620155039,320569.96872013,qk_ar
-12,256,256,512,0.0016183416350267381,3072,805306368,811597824.0,0.9922480620155039,316373.2483416833,qk_ar
-12,256,256,256,0.0008325934893977947,3072,402653184,405798912.0,0.9922480620155039,307472.9784221131,qk_ar
-12,256,256,128,0.0004389725746987952,3072,201326592,202899456.0,0.9922480620155039,291589.9702568624,qk_ar
-12,256,256,64,0.00024191767449664432,3072,100663296,101449728.0,0.9922480620155039,264552.8076159138,qk_ar
-12,256,256,32,0.0001431546143572621,3072,50331648,50724864.0,0.9922480620155039,223534.53392804778,qk_ar
-12,256,256,16,9.404283597678917e-05,3072,25165824,25362432.0,0.9922480620155039,170135.23501087292,qk_ar
-12,256,256,8,6.855550037091989e-05,3072,12582912,12681216.0,0.9922480620155039,116693.773026467,qk_ar
-12,256,256,4,5.4802094978165945e-05,3072,6291456,6340608.0,0.9922480620155039,72989.91036006316,qk_ar
-12,256,256,2,4.608510707869206e-05,3072,3145728,3170304.0,0.9922480620155039,43397.96795057727,qk_ar

research/micro/embedd_micro.py

@@ -1,594 +0,0 @@
-# python embedd_micro.py --use_int8 Fastest
-
-import argparse
-import time
-from dataclasses import dataclass
-from typing import Dict, List, Optional, Tuple
-
-import numpy as np
-import torch
-from torch import nn
-from torchao import quantize_
-from transformers import AutoModel, BitsAndBytesConfig
-from tqdm import tqdm
-from contextlib import contextmanager
-
-@dataclass
-class BenchmarkConfig:
-    model_path: str
-    batch_sizes: List[int]
-    seq_length: int
-    num_runs: int
-    use_fp16: bool = True
-    use_int4: bool = False
-    use_int8: bool = False  # Add this parameter
-    use_cuda_graphs: bool = False
-    use_flash_attention: bool = False
-    use_linear8bitlt: bool = False
-
-
-class CUDAGraphContainer:
-    """Container for managing CUDA graphs for different batch sizes."""
-    
-    def __init__(self, model: nn.Module, seq_length: int):
-        self.model = model
-        self.seq_length = seq_length
-        self.graphs: Dict[int, CUDAGraphWrapper] = {}
-    
-    def get_or_create(self, batch_size: int) -> 'CUDAGraphWrapper':
-        if batch_size not in self.graphs:
-            self.graphs[batch_size] = CUDAGraphWrapper(
-                self.model, batch_size, self.seq_length
-            )
-        return self.graphs[batch_size]
-
-
-class CUDAGraphWrapper:
-    """Wrapper for CUDA graph capture and replay."""
-    
-    def __init__(self, model: nn.Module, batch_size: int, seq_length: int):
-        self.model = model
-        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()
-        
-        # 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
-            )
-    
-    def _create_random_batch(self, batch_size: int, seq_length: int) -> torch.Tensor:
-        return torch.randint(
-            0, 1000, (batch_size, seq_length), 
-            device="cuda", 
-            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
-                )
-    
-    def __call__(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
-        self.static_input.copy_(input_ids)
-        self.static_attention_mask.copy_(attention_mask)
-        self.graph.replay()
-        return self.static_output
-
-
-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
-        model = model.cuda()
-        print("- Model moved to GPU")
-        
-        # 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
-        if 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
-        if config.use_flash_attention:
-            try:
-                from flash_attn.flash_attention import FlashAttention
-                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
-        try:
-            from xformers.ops import 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")
-        
-        model.eval()
-        print("- Model set to eval mode")
-        
-        return model
-
-
-class Timer:
-    """Handles accurate GPU timing using CUDA events."""
-    
-    def __init__(self):
-        self.start_event = torch.cuda.Event(enable_timing=True)
-        self.end_event = torch.cuda.Event(enable_timing=True)
-    
-    @contextmanager
-    def timing(self):
-        self.start_event.record()
-        yield
-        self.end_event.record()
-        self.end_event.synchronize()
-    
-    def elapsed_time(self) -> float:
-        return self.start_event.elapsed_time(self.end_event) / 1000  # ms to seconds
-
-
-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")
-            
-            self.cuda_graphs = (
-                CUDAGraphContainer(self.model, config.seq_length)
-                if config.use_cuda_graphs
-                else None
-            )
-            self.timer = Timer()
-        except Exception as e:
-            print(f"ERROR in benchmark initialization: {str(e)}")
-            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:
-                    print("- Using custom Linear8bitLt replacement for all linear layers")
-                    
-                    # 加载原始模型（不使用量化配置）
-                    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
-                    model = AutoModel.from_pretrained(
-                        self.config.model_path,
-                        torch_dtype=compute_dtype,
-                    )
-                    
-                    # 定义替换函数
-                    def replace_linear_with_linear8bitlt(model):
-                        """递归地将模型中的所有nn.Linear层替换为Linear8bitLt"""
-                        for name, module in list(model.named_children()):
-                            if isinstance(module, nn.Linear):
-                                # 获取原始线性层的参数
-                                in_features = module.in_features
-                                out_features = module.out_features
-                                bias = module.bias is not None
-                                
-                                # 创建8bit线性层
-                                # 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
-                                )
-                                
-                                # 复制权重和偏置
-                                new_module.weight.data = module.weight.data
-                                if bias:
-                                    new_module.bias.data = module.bias.data
-                                    
-                                # 替换模块
-                                setattr(model, name, new_module)
-                            else:
-                                # 递归处理子模块
-                                replace_linear_with_linear8bitlt(module)
-                        
-                        return model
-                    
-                    # 替换所有线性层
-                    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 "cpu"
-                    model = model.to(device)
-                    
-                    print("- All linear layers replaced with Linear8bitLt")
-                    
-                else:
-                    # 使用原来的Int4量化方法
-                    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"
-                    )
-                    
-                    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
-                    )
-                
-                # 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:
-                    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.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
-                try:
-                    from xformers.ops import 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")
-            # Int8 quantization using HuggingFace integration
-            # Int8 quantization using TorchAO
-            elif self.config.use_int8:
-                print("- Using TorchAO for Int8 dynamic activation and Int8 weight quantization")
-                
-                # Import the quantize_ function and the quantization config
-                from torchao.quantization import quantize_, int8_dynamic_activation_int8_weight
-                print("- Successfully imported TorchAO")
-                
-                # Load model normally first
-                # set default to half
-                import torch
-                torch.set_default_dtype(torch.bfloat16)
-                model = AutoModel.from_pretrained(
-                    self.config.model_path,
-                    device_map="auto"
-                )
-                
-                print("- Model loaded in full precision")
-                print(f"- Model type: {type(model)}")
-                
-                # Apply quantization - call the function to get the config, then apply it
-                # quantize_(model, int8_dynamic_activation_int8_weight())
-                # from torchao.quantization import quantize_, Int8DynamicActivationInt8WeightConfig,int8_dynamic_activation_int8_semi_sparse_weight,int4_weight_only,Int8DynActInt4WeightGPTQQuantizer,int8_dynamic_activation_int4_weight,Int8DynamicActivationInt4WeightConfig,Int4DynamicActivationInt4WeightConfig
-                from torchao.quantization import quantize_, Int8DynamicActivationInt8WeightConfig
-                quantize_(model, Int8DynamicActivationInt8WeightConfig())
-                print("- Model successfully quantized with int8 weights and int8 activations")
-                # add torch compile
-                model = torch.compile(model)
-                # For older PyTorch versions that have issues with tensor subclasses
-                from torchao.utils import unwrap_tensor_subclass
-                import torch
-                if hasattr(torch, '_version') and not torch.version >= "2.5.0":
-                    print("- Unwrapping tensor subclasses for compatibility with older PyTorch")
-                    unwrap_tensor_subclass(model)
-                
-                # Apply optimizations
-                if 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")
-                
-                # Set to eval mode
-                model.eval()
-                print("- Model set to eval mode")
-                
-                # For better performance with int8 dynamic quantization
-                torch._inductor.config.force_fuse_int_mm_with_mul = True
-                print("- Enabled fusion of int matmul with mul operations")
-
-                
-                
-            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)}")
-            import traceback
-            traceback.print_exc()
-            raise
-    
-    def _create_random_batch(self, batch_size: int) -> torch.Tensor:
-        return torch.randint(
-            0, 1000,
-            (batch_size, self.config.seq_length),
-            device="cuda",
-            dtype=torch.long
-        )
-    
-    def _run_inference(
-        self,
-        input_ids: torch.Tensor,
-        cuda_graph_wrapper: Optional[CUDAGraphWrapper] = None
-    ) -> Tuple[float, torch.Tensor]:
-        attention_mask = torch.ones_like(input_ids)
-        
-        with torch.no_grad(), self.timer.timing():
-            if cuda_graph_wrapper is not None:
-                output = cuda_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]]:
-        results = {}
-        
-        # Reset peak memory stats
-        torch.cuda.reset_peak_memory_stats()
-        
-        for batch_size in self.config.batch_sizes:
-            print(f"\nTesting batch size: {batch_size}")
-            times = []
-            
-            # Get or create CUDA graph for this batch size
-            cuda_graph_wrapper = (
-                self.cuda_graphs.get_or_create(batch_size)
-                if self.cuda_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:
-                    elapsed_time, output = self._run_inference(input_ids, cuda_graph_wrapper)
-                    if i == 0:  # Only print on first run
-                        print(f"Output shape: {output.last_hidden_state.shape}")
-                    times.append(elapsed_time)
-                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
-        peak_memory_gb = torch.cuda.max_memory_allocated() / (1024 ** 3)
-        print(f"\nPeak GPU memory usage: {peak_memory_gb:.2f} GB")
-        
-        # Add memory info to results
-        for batch_size in results:
-            results[batch_size]["peak_memory_gb"] = peak_memory_gb
-        
-        return results
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Model Inference Benchmark")
-    parser.add_argument(
-        "--model_path",
-        type=str,
-        default="facebook/contriever",
-        help="Path to the model",
-    )
-    parser.add_argument(
-        "--batch_sizes",
-        type=str,
-        default="1,2,4,8,10,16,20,32,40,64,128,256,512,1024,2048,4096,8192",
-        help="Comma-separated list of batch sizes",
-    )
-    parser.add_argument(
-        "--seq_length",
-        type=int,
-        default=256,
-        help="Sequence length for input",
-    )
-    parser.add_argument(
-        "--num_runs",
-        type=int,
-        default=5,
-        help="Number of runs for each batch size",
-    )
-    parser.add_argument(
-        "--use_fp16",
-        action="store_true",
-        help="Enable FP16 inference",
-    )
-    parser.add_argument(
-        "--use_int4",
-        action="store_true",
-        help="Enable INT4 quantization using bitsandbytes",
-    )
-    parser.add_argument(
-        "--use_int8",
-        action="store_true",
-        help="Enable INT8 quantization for both activations and weights using bitsandbytes",
-    )
-    parser.add_argument(
-        "--use_cuda_graphs",
-        action="store_true",
-        help="Enable CUDA Graphs optimization",
-    )
-    parser.add_argument(
-        "--use_flash_attention",
-        action="store_true",
-        help="Enable Flash Attention 2 if available",
-    )
-    parser.add_argument(
-        "--use_linear8bitlt",
-        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(",")],
-        seq_length=args.seq_length,
-        num_runs=args.num_runs,
-        use_fp16=args.use_fp16,
-        use_int4=args.use_int4,
-        use_int8=args.use_int8,  # Add this line
-        use_cuda_graphs=args.use_cuda_graphs,
-        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 "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
-            )
-        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()

research/micro/embedd_micro_seq.py

@@ -1,376 +0,0 @@
-import argparse
-import time
-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
-from tqdm import tqdm
-from contextlib import contextmanager
-import math
-
-@dataclass
-class BenchmarkConfig:
-    model_path: str
-    batch_sizes: List[int]
-    seq_length: int
-    num_runs: int
-    use_fp16: bool = True
-    use_cuda_graphs: bool = False
-    use_flash_attention: bool = False
-    max_batch_size: int = 256  # Maximum batch size before splitting
-
-
-class CUDAGraphContainer:
-    """Container for managing CUDA graphs for different batch sizes."""
-    
-    def __init__(self, model: nn.Module, seq_length: int, max_batch_size: int):
-        self.model = model
-        self.seq_length = seq_length
-        self.max_batch_size = max_batch_size
-        self.graphs: Dict[int, CUDAGraphWrapper] = {}
-    
-    def get_or_create(self, batch_size: int) -> 'CUDAGraphWrapper':
-        # For CUDA graphs, we always use the actual batch size or max_batch_size
-        effective_batch_size = min(batch_size, self.max_batch_size)
-        
-        if effective_batch_size not in self.graphs:
-            self.graphs[effective_batch_size] = CUDAGraphWrapper(
-                self.model, effective_batch_size, self.seq_length
-            )
-        return self.graphs[effective_batch_size]
-
-
-class CUDAGraphWrapper:
-    """Wrapper for CUDA graph capture and replay."""
-    
-    def __init__(self, model: nn.Module, batch_size: int, seq_length: int):
-        self.model = model
-        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()
-        
-        # 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
-            )
-    
-    def _create_random_batch(self, batch_size: int, seq_length: int) -> torch.Tensor:
-        return torch.randint(
-            0, 1000, (batch_size, seq_length), 
-            device="cuda", 
-            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
-                )
-    
-    def __call__(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
-        self.static_input.copy_(input_ids)
-        self.static_attention_mask.copy_(attention_mask)
-        self.graph.replay()
-        return self.static_output
-
-
-class ModelOptimizer:
-    """Applies various optimizations to the model."""
-    
-    @staticmethod
-    def optimize(model: nn.Module, config: BenchmarkConfig) -> nn.Module:
-        print("\nApplying model optimizations:")
-        
-        # Move to GPU
-        model = model.cuda()
-        print("- Model moved to GPU")
-        
-        # FP16
-        if config.use_fp16:
-            model = model.half()
-            print("- Using FP16 precision")
-        
-        # Check if using SDPA
-        if 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)")
-                # No need to do anything as it's automatically enabled
-            else:
-                print("- PyTorch SDPA not available")
-        
-        # Flash Attention
-        if config.use_flash_attention:
-            try:
-                from flash_attn.flash_attention import FlashAttention
-                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")
-        
-        # Optimize LayerNorm
-        try:
-            num_layernorms = 0
-            for module in model.modules():
-                if isinstance(module, torch.nn.LayerNorm):
-                    module.forward = torch.jit.script(module.forward)
-                    num_layernorms += 1
-            if num_layernorms > 0:
-                print(f"- Optimized {num_layernorms} LayerNorm modules with TorchScript")
-        except Exception as e:
-            print(f"- LayerNorm optimization failed: {e}")
-        
-        # Memory efficient attention
-        try:
-            from xformers.ops import memory_efficient_attention
-            model.enable_xformers_memory_efficient_attention()
-            print("- Enabled xformers memory efficient attention")
-        except (ImportError, AttributeError):
-            print("- Xformers not available")
-        
-        model.eval()
-        print("- Model set to eval mode")
-        
-        return model
-
-
-class Timer:
-    """Handles accurate GPU timing using CUDA events."""
-    
-    def __init__(self):
-        self.start_event = torch.cuda.Event(enable_timing=True)
-        self.end_event = torch.cuda.Event(enable_timing=True)
-    
-    @contextmanager
-    def timing(self):
-        self.start_event.record()
-        yield
-        self.end_event.record()
-        self.end_event.synchronize()
-    
-    def elapsed_time(self) -> float:
-        return self.start_event.elapsed_time(self.end_event) / 1000  # ms to seconds
-
-
-class Benchmark:
-    """Main benchmark runner."""
-    
-    def __init__(self, config: BenchmarkConfig):
-        self.config = config
-        self.model = self._load_model()
-        self.cuda_graphs = (
-            CUDAGraphContainer(self.model, config.seq_length, config.max_batch_size)
-            if config.use_cuda_graphs
-            else None
-        )
-        self.timer = Timer()
-    
-    def _load_model(self) -> nn.Module:
-        print(f"Loading model from {self.config.model_path}...")
-        model = AutoModel.from_pretrained(self.config.model_path)
-        return ModelOptimizer.optimize(model, self.config)
-    
-    def _create_random_batch(self, batch_size: int) -> torch.Tensor:
-        return torch.randint(
-            0, 1000,
-            (batch_size, self.config.seq_length),
-            device="cuda",
-            dtype=torch.long
-        )
-    
-    def _run_inference(
-        self,
-        input_ids: torch.Tensor,
-        cuda_graph_wrapper: Optional[CUDAGraphWrapper] = None
-    ) -> Tuple[float, torch.Tensor]:
-        attention_mask = torch.ones_like(input_ids)
-        original_batch_size = input_ids.shape[0]
-        print(f"Original input_ids shape: {input_ids.shape}")
-        
-        # Split large batches to avoid OOM
-        max_batch_size = self.config.max_batch_size
-        if original_batch_size > max_batch_size:
-            print(f"Splitting batch of size {original_batch_size} into chunks of {max_batch_size}")
-            total_time = 0
-            outputs = []
-            
-            with torch.no_grad():
-                for i in range(0, original_batch_size, max_batch_size):
-                    end_idx = min(i + max_batch_size, original_batch_size)
-                    batch_slice = input_ids[i:end_idx]
-                    mask_slice = attention_mask[i:end_idx]
-                    
-                    print(f"Processing chunk {i//max_batch_size + 1}: shape {batch_slice.shape}")
-                    
-                    # Use CUDA graph if available (with the smaller batch size)
-                    chunk_cuda_graph = None
-                    if cuda_graph_wrapper is not None:
-                        chunk_cuda_graph = self.cuda_graphs.get_or_create(batch_slice.shape[0])
-                    
-                    with self.timer.timing():
-                        if chunk_cuda_graph is not None:
-                            chunk_output = chunk_cuda_graph(batch_slice, mask_slice)
-                        else:
-                            chunk_output = self.model(input_ids=batch_slice, attention_mask=mask_slice)
-                    
-                    total_time += self.timer.elapsed_time()
-                    outputs.append(chunk_output.last_hidden_state)
-                
-                # Combine outputs
-                combined_output = torch.cat(outputs, dim=0)
-                print(f"Combined output shape: {combined_output.shape}")
-                
-                # Create a wrapper object similar to model output to maintain consistency
-                class DummyOutput:
-                    def __init__(self, hidden_states):
-                        self.last_hidden_state = hidden_states
-                
-                output = DummyOutput(combined_output)
-                return total_time, output
-        else:
-            # Process normally for small batches
-            with torch.no_grad(), self.timer.timing():
-                if cuda_graph_wrapper is not None:
-                    output = cuda_graph_wrapper(input_ids, attention_mask)
-                else:
-                    output = self.model(input_ids=input_ids, attention_mask=attention_mask)
-            
-            print(f"Output shape: {output.last_hidden_state.shape}")
-            return self.timer.elapsed_time(), output
-    
-    def run(self) -> Dict[int, Dict[str, float]]:
-        results = {}
-        
-        for batch_size in self.config.batch_sizes:
-            print(f"\nTesting batch size: {batch_size}")
-            times = []
-            
-            # Get or create CUDA graph for this batch size
-            cuda_graph_wrapper = None
-            if self.cuda_graphs is not None:
-                if batch_size <= self.config.max_batch_size:
-                    cuda_graph_wrapper = self.cuda_graphs.get_or_create(batch_size)
-                else:
-                    # For large batches, we'll use the max_batch_size graph in chunks
-                    cuda_graph_wrapper = True  # Just a flag to indicate we want to use CUDA graphs
-            
-            # Pre-allocate input tensor
-            input_ids = self._create_random_batch(batch_size)
-            
-            # Run benchmark
-            for run_idx in tqdm(range(self.config.num_runs), desc=f"Batch size {batch_size}"):
-                elapsed_time, _ = self._run_inference(input_ids, cuda_graph_wrapper)
-                times.append(elapsed_time)
-                print(f"Run {run_idx+1}: {elapsed_time:.4f}s")
-            
-            # 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")
-        
-        return results
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Model Inference Benchmark")
-    parser.add_argument(
-        "--model_path",
-        type=str,
-        default="facebook/contriever",
-        help="Path to the model",
-    )
-    parser.add_argument(
-        "--batch_sizes",
-        type=str,
-        default="1,2,4,8,16,32,64,128,256,512,1024,2048,4096",
-        help="Comma-separated list of batch sizes",
-    )
-    parser.add_argument(
-        "--seq_length",
-        type=int,
-        default=256,
-        help="Sequence length for input",
-    )
-    parser.add_argument(
-        "--num_runs",
-        type=int,
-        default=5,
-        help="Number of runs for each batch size",
-    )
-    parser.add_argument(
-        "--no_fp16",
-        action="store_true",
-        help="Disable FP16 inference",
-    )
-    parser.add_argument(
-        "--use_cuda_graphs",
-        action="store_true",
-        help="Enable CUDA Graphs optimization",
-    )
-    parser.add_argument(
-        "--use_flash_attention",
-        action="store_true",
-        help="Enable Flash Attention 2 if available",
-    )
-    parser.add_argument(
-        "--max_batch_size",
-        type=int,
-        default=256,
-        help="Maximum batch size before splitting to prevent OOM",
-    )
-    
-    args = parser.parse_args()
-    
-    config = BenchmarkConfig(
-        model_path=args.model_path,
-        batch_sizes=[int(bs) for bs in args.batch_sizes.split(",")],
-        seq_length=args.seq_length,
-        num_runs=args.num_runs,
-        use_fp16=not args.no_fp16,
-        use_cuda_graphs=args.use_cuda_graphs,
-        use_flash_attention=args.use_flash_attention,
-        max_batch_size=args.max_batch_size,
-    )
-    
-    benchmark = Benchmark(config)
-    results = benchmark.run()
-    
-    # Print overall summary
-    print("\n===== BENCHMARK SUMMARY =====")
-    print(f"Model: {config.model_path}")
-    print(f"Sequence Length: {config.seq_length}")
-    print(f"FP16: {config.use_fp16}")
-    print(f"CUDA Graphs: {config.use_cuda_graphs}")
-    print(f"Flash Attention: {config.use_flash_attention}")
-    print(f"Max Batch Size: {config.max_batch_size}")
-    print("\nResults:")
-    
-    print("\nBatch Size | Avg Time (s) | Throughput (seq/s)")
-    print("-" * 50)
-    for bs in sorted(results.keys()):
-        r = results[bs]
-        print(f"{bs:^10} | {r['avg_time']:^12.4f} | {r['throughput']:^17.2f}")
-
-
-if __name__ == "__main__":
-    main()

research/micro/int4benchmark.py

@@ -1,218 +0,0 @@
-import torch
-import torch.nn as nn
-import time
-import torch.nn.functional as F
-
-# Import necessary functions from the quantize.py file
-def get_group_qparams(w, n_bit=4, groupsize=128):
-    # needed for GPTQ with padding
-    if groupsize > w.shape[-1]:
-        groupsize = w.shape[-1]
-    assert groupsize > 1
-    assert w.shape[-1] % groupsize == 0
-    assert w.dim() == 2
-
-    to_quant = w.reshape(-1, groupsize)
-    assert torch.isnan(to_quant).sum() == 0
-
-    max_val = to_quant.amax(dim=1, keepdim=True)
-    min_val = to_quant.amin(dim=1, keepdim=True)
-    max_int = 2**n_bit - 1
-    scales = (max_val - min_val).clamp(min=1e-6) / max_int
-    zeros = min_val + scales * (2 ** (n_bit - 1))
-    return scales.to(torch.bfloat16).reshape(w.shape[0], -1), zeros.to(
-        torch.bfloat16
-    ).reshape(w.shape[0], -1)
-
-def pack_scales_and_zeros(scales, zeros):
-    assert scales.shape == zeros.shape
-    assert scales.dtype == torch.bfloat16
-    assert zeros.dtype == torch.bfloat16
-    return (
-        torch.cat(
-            [
-                scales.reshape(scales.size(0), scales.size(1), 1),
-                zeros.reshape(zeros.size(0), zeros.size(1), 1),
-            ],
-            2,
-        )
-        .transpose(0, 1)
-        .contiguous()
-    )
-
-def group_quantize_tensor(w, n_bit=4, groupsize=128):
-    scales, zeros = get_group_qparams(w, n_bit, groupsize)
-    w_int32 = group_quantize_tensor_from_qparams(w, scales, zeros, n_bit, groupsize)
-    scales_and_zeros = pack_scales_and_zeros(scales, zeros)
-    return w_int32, scales_and_zeros
-
-def group_quantize_tensor_from_qparams(w, scales, zeros, n_bit=4, groupsize=128):
-    assert groupsize > 1
-    # needed for GPTQ single column quantize
-    if groupsize > w.shape[-1] and scales.shape[-1] == 1:
-        groupsize = w.shape[-1]
-
-    assert w.shape[-1] % groupsize == 0
-    assert w.dim() == 2
-
-    to_quant = w.reshape(-1, groupsize)
-    assert torch.isnan(to_quant).sum() == 0
-
-    scales = scales.reshape(-1, 1)
-    zeros = zeros.reshape(-1, 1)
-    min_val = zeros - scales * (2 ** (n_bit - 1))
-    max_int = 2**n_bit - 1
-    min_int = 0
-    w_int32 = (
-        to_quant.sub(min_val)
-        .div(scales)
-        .round()
-        .clamp_(min_int, max_int)
-        .to(torch.int32)
-        .reshape_as(w)
-    )
-
-    return w_int32
-
-def prepare_int4_weight_and_scales_and_zeros(weight_bf16, groupsize, inner_k_tiles):
-    weight_int32, scales_and_zeros = group_quantize_tensor(
-        weight_bf16, n_bit=4, groupsize=groupsize
-    )
-    weight_int4pack = torch.ops.aten._convert_weight_to_int4pack(weight_int32, inner_k_tiles)
-    return weight_int4pack, scales_and_zeros
-
-def linear_forward_int4(x, weight_int4pack, scales_and_zeros, out_features, groupsize):
-    origin_x_size = x.size()
-    x = x.reshape(-1, origin_x_size[-1])
-    c = torch.ops.aten._weight_int4pack_mm(x, weight_int4pack, groupsize, scales_and_zeros)
-    new_shape = origin_x_size[:-1] + (out_features,)
-    c = c.reshape(new_shape)
-    return c
-
-class WeightOnlyInt4Linear(torch.nn.Module):
-    __constants__ = ['in_features', 'out_features']
-    in_features: int
-    out_features: int
-    weight: torch.Tensor
-
-    def __init__(
-            self, in_features: int, out_features: int,
-            bias=False, device=None, dtype=None, groupsize: int = 128, inner_k_tiles: int = 8
-    ) -> None:
-        super().__init__()
-        self.in_features = in_features
-        self.out_features = out_features
-        self.groupsize = groupsize
-        self.inner_k_tiles = inner_k_tiles
-
-        assert out_features % 8 == 0, "require out_features % 8 == 0"
-        assert in_features % (inner_k_tiles * 16) == 0, "require in_features % (innerKTiles * 16) == 0"
-        self.register_buffer(
-            "weight",
-            torch.empty((out_features // 8, in_features // (inner_k_tiles * 16), 32, inner_k_tiles // 2), dtype=torch.int32)
-        )
-        self.register_buffer(
-            "scales_and_zeros",
-            torch.empty((in_features // groupsize, out_features, 2), dtype=torch.bfloat16)
-        )
-
-    def forward(self, input: torch.Tensor) -> torch.Tensor:
-        input = input.to(torch.bfloat16)
-        return linear_forward_int4(
-            input,
-            self.weight, self.scales_and_zeros, self.out_features, self.groupsize
-        )
-
-# Define dimensions that satisfy the requirements for INT4 quantization
-# in_features must be divisible by inner_k_tiles * 16
-# out_features must be divisible by 8
-in_features = 1024  # Must be divisible by inner_k_tiles * 16
-out_features = 2048  # Must be divisible by 8
-groupsize = 128
-inner_k_tiles = 8
-
-# Create models
-fp16_model = nn.Sequential(
-    nn.Linear(in_features, out_features, bias=False)
-)
-
-# Create INT4 model
-int4_model = nn.Sequential(
-    WeightOnlyInt4Linear(in_features, out_features, bias=False, 
-                         groupsize=groupsize, inner_k_tiles=inner_k_tiles)
-)
-
-# Quantize the weights and set up the INT4 model
-with torch.no_grad():
-    # Convert FP16 weights to INT4
-    fp16_weight = fp16_model[0].weight.data.to(torch.bfloat16)
-    weight_int4pack, scales_and_zeros = prepare_int4_weight_and_scales_and_zeros(
-        fp16_weight, groupsize, inner_k_tiles
-    )
-    
-    # Set the quantized weights in the INT4 model
-    int4_model[0].weight.copy_(weight_int4pack)
-    int4_model[0].scales_and_zeros.copy_(scales_and_zeros)
-
-# Move models to GPU
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-fp16_model = fp16_model.to(device)
-int4_model = int4_model.to(device)
-
-# Create random input tensor
-batch_size = 1024
-input_tensor = torch.randn(batch_size, in_features, device=device)
-input_tensor_bf16 = input_tensor.to(torch.bfloat16)
-
-# Speed test function
-def speed_test(model, input_tensor, name, num_iterations=100):
-    # Warmup
-    for _ in range(10):
-        _ = model(input_tensor)
-    
-    # Actual timing
-    torch.cuda.synchronize()
-    start_time = time.time()
-    
-    for _ in range(num_iterations):
-        _ = model(input_tensor)
-    
-    torch.cuda.synchronize()
-    end_time = time.time()
-    
-    avg_time = (end_time - start_time) / num_iterations
-    print(f"{name} model: {avg_time:.6f} seconds per iteration")
-    return avg_time
-
-# Run speed tests
-with torch.no_grad():  # Disable gradient calculation for inference
-    print(f"Running benchmark with batch_size={batch_size}, in_features={in_features}, out_features={out_features}")
-    print(f"INT4 parameters: groupsize={groupsize}, inner_k_tiles={inner_k_tiles}")
-    
-    fp16_time = speed_test(fp16_model, input_tensor_bf16, "FP16")
-    int4_time = speed_test(int4_model, input_tensor, "INT4")
-    
-    # Calculate speedup
-    speedup = fp16_time / int4_time
-    print(f"INT4 is {speedup:.2f}x faster than FP16")
-    
-    # Calculate memory savings
-    fp16_memory = fp16_model[0].weight.nelement() * fp16_model[0].weight.element_size()
-    int4_memory = (int4_model[0].weight.nelement() * int4_model[0].weight.element_size() + 
-                  int4_model[0].scales_and_zeros.nelement() * int4_model[0].scales_and_zeros.element_size())
-    
-    memory_reduction = fp16_memory / int4_memory
-    print(f"Memory reduction: {memory_reduction:.2f}x ({fp16_memory/1024/1024:.2f} MB vs {int4_memory/1024/1024:.2f} MB)")
-    
-    # Check accuracy
-    with torch.no_grad():
-        fp16_output = fp16_model(input_tensor_bf16)
-        int4_output = int4_model(input_tensor)
-        
-        # Calculate error metrics
-        abs_error = torch.abs(fp16_output - int4_output)
-        rel_error = abs_error / (torch.abs(fp16_output) + 1e-7)
-        
-        print(f"Mean absolute error: {abs_error.mean().item():.6f}")
-        print(f"Max absolute error: {abs_error.max().item():.6f}")
-        print(f"Mean relative error: {rel_error.mean().item():.6f}") 

research/micro/int8.py

@@ -1,83 +0,0 @@
-import torch
-import nvmath.bindings.cublas
-import ctypes
-
-# 创建 CUBLAS 句柄
-handle = nvmath.bindings.cublas.create()
-
-# 准备数据 - 使用 uint8 类型，并确保内存连续
-m, n, k = 64, 32, 48
-a = (torch.rand(m, k, device="cuda") * 255).to(torch.uint8).contiguous()
-b = (torch.rand(k, n, device="cuda") * 255).to(torch.uint8).contiguous()
-c = torch.zeros(m, n, device="cuda", dtype=torch.uint8).contiguous()
-
-# 确保张量在 CUDA 上
-assert a.is_cuda and b.is_cuda and c.is_cuda
-# 确保张量是连续的
-assert a.is_contiguous() and b.is_contiguous() and c.is_contiguous()
-
-# 获取指针
-a_ptr = a.data_ptr()
-b_ptr = b.data_ptr()
-c_ptr = c.data_ptr()
-
-# 设置参数
-transa = 0  # CUBLAS_OP_N (不转置)
-transb = 0  # CUBLAS_OP_N (不转置)
-transc = 0  # CUBLAS_OP_N (不转置)
-
-# 设置偏置值
-a_bias = 0
-b_bias = 0
-c_bias = 0
-
-# 设置正确的 leading dimensions
-lda = k  # A 的 leading dimension
-ldb = n  # B 的 leading dimension
-ldc = n  # C 的 leading dimension
-
-c_mult = 1
-c_shift = 0
-
-# 打印调试信息
-print(f"a shape: {a.shape}, a_ptr: {a_ptr}")
-print(f"b shape: {b.shape}, b_ptr: {b_ptr}")
-print(f"c shape: {c.shape}, c_ptr: {c_ptr}")
-
-try:
-    # 调用 uint8gemm_bias
-    nvmath.bindings.cublas.uint8gemm_bias(
-        handle,
-        transa, transb, transc,
-        m, n, k,
-        a_ptr, a_bias, lda,
-        b_ptr, b_bias, ldb,
-        c_ptr, c_bias, ldc,
-        c_mult, c_shift
-    )
-except Exception as e:
-    print(f"Error: {e}")
-    # 尝试使用 ctypes 转换指针
-    a_ptr_c = ctypes.c_void_p(a_ptr).value
-    b_ptr_c = ctypes.c_void_p(b_ptr).value
-    c_ptr_c = ctypes.c_void_p(c_ptr).value
-    
-    print(f"Using ctypes: a_ptr: {a_ptr_c}, b_ptr: {b_ptr_c}, c_ptr: {c_ptr_c}")
-    
-    # 再次尝试调用
-    nvmath.bindings.cublas.uint8gemm_bias(
-        handle,
-        transa, transb, transc,
-        m, n, k,
-        a_ptr_c, a_bias, lda,
-        b_ptr_c, b_bias, ldb,
-        c_ptr_c, c_bias, ldc,
-        c_mult, c_shift
-    )
-
-# 销毁 CUBLAS 句柄
-nvmath.bindings.cublas.destroy(handle)
-
-# 打印结果
-print("Result:")
-print(c)

research/micro/llm_compress.py

@@ -1,23 +0,0 @@
-from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
-from llmcompressor.modifiers.quantization import GPTQModifier
-from llmcompressor import oneshot
-
-# Select quantization algorithm. In this case, we:
-#   * apply SmoothQuant to make the activations easier to quantize
-#   * quantize the weights to int8 with GPTQ (static per channel)
-#   * quantize the activations to int8 (dynamic per token)
-recipe = [
-    SmoothQuantModifier(smoothing_strength=0.8),
-    GPTQModifier(scheme="W8A8", targets="Linear", ignore=["lm_head"]),
-]
-
-# Apply quantization using the built in open_platypus dataset.
-#   * See examples for demos showing how to pass a custom calibration set
-oneshot(
-    model="facebook/contriever",
-    dataset="open_platypus",
-    recipe=recipe,
-    output_dir="contriever-INT4",
-    max_seq_length=2048,
-    num_calibration_samples=512,
-)

research/micro/nvmath_test.py

@@ -1,41 +0,0 @@
-# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
-#
-# SPDX-License-Identifier: Apache-2.0
-
-"""
-This example demonstrates basic matrix multiplication of FP8 tensors.
-
-In narrow-precision operations, quantization scales must be provided for each tensor. These
-scales are used to dequantize input operands and quantize the result. Without proper
-scaling, the results of FP8 operations will likely exceed the type's range.
-
-FP8 is only supported with cuBLAS 12.8 or newer and on devices with compute
-capability 8.9 or higher.
-"""
-
-import torch
-
-import nvmath
-
-# Prepare sample input data. Note that N, M and K must be divisible by 16 for FP8.
-# cuBLAS requires B to be column-major, so we first create a row-major tensor and then
-# transpose it.
-m, n, k = 64, 32, 48
-a = (torch.rand(m, k, device="cuda") * 10).type(torch.float8_e4m3fn)
-b = (torch.rand(n, k, device="cuda") * 10).type(torch.float8_e4m3fn).T
-
-# Prepare quantization scales. The scales must allow the result to fit within the dynamic
-# range of the data type used. Scales can be provided either as a dictionary or as a
-# MatmulQuantizationScales object. Note that scales are only allowed for FP8 operands.
-scales = {"a": 1, "b": 1, "d": 0.1}
-
-# Perform the multiplication. The result of the multiplication will be:
-# (scales.a * A) @ (scales.b * B) * scales.d
-result = nvmath.linalg.advanced.matmul(a, b, quantization_scales=scales)
-
-# Check how scaling helped to fit into the dynamic range of float8_e4m3fn type.
-result_without_scaling = nvmath.linalg.advanced.matmul(a, b, quantization_scales={"a": 1, "b": 1, "d": 1})
-print("Without scaling, most of the elements were clamped to the maximum value of float8_e4m3fn type (448):")
-print(result_without_scaling)
-print(f"\nWith D scale set to {scales['d']}, they were scaled down to fit into the dynamic range of float8_e4m3fn:")
-print(result)

research/micro/save_small_model.py

@@ -1,58 +0,0 @@
-import os
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from pathlib import Path
-
-def save_model_in_pth_format(model_name, output_dir):
-    """
-    Download a model from Hugging Face and save it in PTH format
-    for use with quantization benchmarks.
-    
-    Args:
-        model_name: Name of the model on Hugging Face
-        output_dir: Directory to save the model
-    """
-    print(f"Loading model {model_name}...")
-    
-    # Create output directory if it doesn't exist
-    os.makedirs(output_dir, exist_ok=True)
-    
-    # Load tokenizer and model
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    model = AutoModelForCausalLM.from_pretrained(
-        model_name,
-        torch_dtype=torch.bfloat16,
-        low_cpu_mem_usage=True
-    )
-    
-    # Save tokenizer
-    tokenizer.save_pretrained(output_dir)
-    
-    # Extract and save the model weights in PTH format
-    model_state_dict = model.state_dict()
-    
-    # Save the model weights
-    model_path = Path(output_dir) / "model.pth"
-    torch.save(model_state_dict, model_path)
-    
-    print(f"Model saved to {model_path}")
-    
-    # Print model size information
-    param_count = sum(p.numel() for p in model.parameters())
-    model_size_mb = sum(p.numel() * p.element_size() for p in model.parameters()) / (1024 * 1024)
-    
-    print(f"Model parameters: {param_count:,}")
-    print(f"Model size: {model_size_mb:.2f} MB")
-    
-    return model_path
-
-if __name__ == "__main__":
-    # Use a small model for testing
-    model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
-    output_dir = "./tinyllama-1.1b-chat"
-    
-    model_path = save_model_in_pth_format(model_name, output_dir)
-    
-    print("\nYou can now use this model with the INT4 benchmark script.")
-    print("Example command:")
-    print(f"python int4benchmark.py --model_path {model_path}") 

research/micro/transformer-batching-benchmark.ipynb

@@ -1,677 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "id": "cab91cfc",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/home/ubuntu/Power-RAG/.venv/lib/python3.10/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"
-     ]
-    }
-   ],
-   "source": [
-    "import copy\n",
-    "import dataclasses\n",
-    "import os\n",
-    "import time\n",
-    "import pathlib\n",
-    "import itertools\n",
-    "import multiprocessing\n",
-    "import scipy\n",
-    "import numpy as np\n",
-    "import pandas as pd\n",
-    "import pickle\n",
-    "import gzip\n",
-    "import threading\n",
-    "import queue\n",
-    "import pytz\n",
-    "import traceback\n",
-    "from datetime import datetime\n",
-    "from tqdm.auto import tqdm, trange\n",
-    "from typing import Any\n",
-    "\n",
-    "import matplotlib.pyplot as plt\n",
-    "import matplotlib.ticker as mtick\n",
-    "%matplotlib inline\n",
-    "%config InlineBackend.figure_format='retina'"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "id": "8d24fbd7",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Sat Apr 12 00:10:05 2025       \n",
-      "+-----------------------------------------------------------------------------------------+\n",
-      "| NVIDIA-SMI 550.120                Driver Version: 550.120        CUDA Version: 12.4     |\n",
-      "|-----------------------------------------+------------------------+----------------------+\n",
-      "| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |\n",
-      "| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |\n",
-      "|                                         |                        |               MIG M. |\n",
-      "|=========================================+========================+======================|\n",
-      "|   0  NVIDIA A10G                    Off |   00000000:00:1E.0 Off |                    0 |\n",
-      "|  0%   27C    P8             15W /  300W |       4MiB /  23028MiB |      0%      Default |\n",
-      "|                                         |                        |                  N/A |\n",
-      "+-----------------------------------------+------------------------+----------------------+\n",
-      "                                                                                         \n",
-      "+-----------------------------------------------------------------------------------------+\n",
-      "| Processes:                                                                              |\n",
-      "|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |\n",
-      "|        ID   ID                                                               Usage      |\n",
-      "|=========================================================================================|\n",
-      "|  No running processes found                                                             |\n",
-      "+-----------------------------------------------------------------------------------------+\n"
-     ]
-    }
-   ],
-   "source": [
-    "!nvidia-smi"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "id": "538b2c11",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def benchmark(f, *, f_setup=None, min_repeat: int, min_secs: float, tqdm_kwargs: dict | None=None) -> np.ndarray:\n",
-    "    latency = []\n",
-    "    \n",
-    "    # First run, ignore min_secs\n",
-    "    if f_setup is not None:\n",
-    "        f_setup()\n",
-    "    st = time.perf_counter_ns()\n",
-    "    f()\n",
-    "    ed = time.perf_counter_ns()\n",
-    "    latency.append((ed-st)/1e9)\n",
-    "    \n",
-    "    # Subsequent runs, until reaching both min_repeat and min_secs\n",
-    "    min_nanos = int(min_secs * 1e9)\n",
-    "    start_nanos = time.perf_counter_ns()\n",
-    "    while True:\n",
-    "        now_nanos = time.perf_counter_ns()\n",
-    "        if len(latency) > min_repeat and now_nanos - start_nanos > min_nanos:\n",
-    "            break\n",
-    "        if f_setup is not None:\n",
-    "            f_setup()\n",
-    "        st = time.perf_counter_ns()\n",
-    "        f()\n",
-    "        ed = time.perf_counter_ns()\n",
-    "        latency.append((ed-st)/1e9)\n",
-    "    return np.array(latency)\n",
-    "\n",
-    "def tail_mean(xs, skip=0.2):\n",
-    "    return xs[int(len(xs) * skip):].mean()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "id": "02c9c9b1",
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "<torch.autograd.grad_mode.set_grad_enabled at 0x7c5afc12b850>"
-      ]
-     },
-     "execution_count": 4,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "import torch\n",
-    "torch.set_grad_enabled(False)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "id": "3405fdc7",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "nd_list = list(itertools.chain(itertools.product([12, 3], [256])))\n",
-    "seqlen_list = [256]\n",
-    "bs_list = [2,4,8,16,32,64,128,256,512,1024,2048]"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "id": "10dc981a",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "[(12, 256), (3, 256)]\n",
-      "[256]\n",
-      "[2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048]\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(nd_list)\n",
-    "print(seqlen_list)\n",
-    "print(bs_list)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "id": "7e0ee385",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def benchmark_dense(out, nd_list, seqlen_list, bs_list):\n",
-    "    seqlen_list = [1] + seqlen_list\n",
-    "    total = len(list(itertools.product(nd_list, seqlen_list, bs_list)))\n",
-    "    pbar = tqdm(total=total)\n",
-    "    for (n, d), seqlen in reversed(list(itertools.product(nd_list, seqlen_list))):\n",
-    "        h = n * d\n",
-    "        maxbs = max(bs_list)\n",
-    "        print(maxbs, n, d, seqlen)\n",
-    "        cache = torch.empty(int(256e6 // 4), dtype=torch.int, device=\"cuda:0\")\n",
-    "        X = torch.rand((maxbs, seqlen, h), dtype=torch.bfloat16, device=\"cuda:0\")\n",
-    "        W = torch.rand((h, h), dtype=torch.bfloat16, device=\"cuda:0\")\n",
-    "        torch.cuda.synchronize()\n",
-    "        for bs in reversed(bs_list):\n",
-    "            pbar.set_postfix(n=n, h=h, d=d, seqlen=seqlen, bs=bs)\n",
-    "            def run():\n",
-    "                torch.matmul(X[:bs], W)\n",
-    "                torch.cuda.synchronize()\n",
-    "            def clear_cache():\n",
-    "                cache.zero_()\n",
-    "                torch.cuda.synchronize()\n",
-    "            latency = benchmark(run, f_setup=clear_cache, min_repeat=20, min_secs=2)\n",
-    "            l = tail_mean(latency)\n",
-    "            out.append({\n",
-    "                \"n\": n,\n",
-    "                \"d\": d,\n",
-    "                \"seqlen\": seqlen,\n",
-    "                \"bs\": bs,\n",
-    "                \"latency\": l\n",
-    "            })\n",
-    "            pbar.update()\n",
-    "        del cache, X, W\n",
-    "        torch.cuda.empty_cache()\n",
-    "    pbar.close()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "id": "c206a502",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def benchmark_qk_init(out, nd_list, seqlen_list, bs_list):\n",
-    "    total = len(list(itertools.product(nd_list, seqlen_list, bs_list)))\n",
-    "    pbar = tqdm(total=total)\n",
-    "    for (n, d), seqlen in reversed(list(itertools.product(nd_list, seqlen_list))):\n",
-    "        h = n * d\n",
-    "        try:\n",
-    "            maxbs = max(b for b in bs_list if b*n*seqlen*d*2*2+b*n*seqlen**2*2 < 80e9)\n",
-    "        except ValueError:\n",
-    "            pbar.update(len(bs_list))\n",
-    "            continue\n",
-    "        cache = torch.empty(int(256e6 // 4), dtype=torch.int, device=\"cuda:0\")\n",
-    "        Qmax = torch.rand((maxbs, n, seqlen, d), dtype=torch.bfloat16, device=\"cuda:0\")\n",
-    "        Kmax = torch.rand((maxbs, n, seqlen, d), dtype=torch.bfloat16, device=\"cuda:0\")\n",
-    "        torch.cuda.synchronize()\n",
-    "        for bs in reversed(bs_list):\n",
-    "            pbar.set_postfix(n=n, h=h, d=d, seqlen=seqlen, bs=bs)\n",
-    "            if bs > maxbs:\n",
-    "                pbar.update()\n",
-    "                continue\n",
-    "            Q = Qmax[:bs]\n",
-    "            K = Kmax[:bs]\n",
-    "            def run():\n",
-    "                torch.bmm(Q.view(bs * n, seqlen, d), K.view(bs * n, seqlen, d).transpose(1, 2))\n",
-    "                torch.cuda.synchronize()\n",
-    "            def clear_cache():\n",
-    "                cache.zero_()\n",
-    "                torch.cuda.synchronize()\n",
-    "            latency = benchmark(run, f_setup=clear_cache, min_repeat=20, min_secs=2)\n",
-    "            l = tail_mean(latency)\n",
-    "            out.append({\n",
-    "                \"n\": n,\n",
-    "                \"d\": d,\n",
-    "                \"seqlen\": seqlen,\n",
-    "                \"bs\": bs,\n",
-    "                \"latency\": l\n",
-    "            })\n",
-    "            pbar.update()\n",
-    "        del cache, Q, K, Qmax, Kmax\n",
-    "        torch.cuda.empty_cache()\n",
-    "    pbar.close()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "id": "a3a2103c",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def benchmark_qk_ar(out, nd_list, seqlen_list, bs_list):\n",
-    "    total = len(list(itertools.product(nd_list, seqlen_list, bs_list)))\n",
-    "    pbar = tqdm(total=total)\n",
-    "    for (n, d), seqlen in reversed(list(itertools.product(nd_list, seqlen_list))):\n",
-    "        h = n * d\n",
-    "        try:\n",
-    "            maxbs = max(b for b in bs_list if b*n*(1+seqlen)*d*2+b*n*seqlen*2 < 80e9)\n",
-    "        except ValueError:\n",
-    "            pbar.update(len(bs_list))\n",
-    "            continue\n",
-    "        cache = torch.empty(int(256e6 // 4), dtype=torch.int, device=\"cuda:0\")\n",
-    "        Qmax = torch.rand((maxbs, n, 1, d), dtype=torch.bfloat16, device=\"cuda:0\")\n",
-    "        Kmax = torch.rand((maxbs, n, seqlen, d), dtype=torch.bfloat16, device=\"cuda:0\")\n",
-    "        torch.cuda.synchronize()\n",
-    "        for bs in reversed(bs_list):\n",
-    "            pbar.set_postfix(n=n, h=h, d=d, seqlen=seqlen, bs=bs)\n",
-    "            if bs > maxbs:\n",
-    "                pbar.update()\n",
-    "                continue\n",
-    "            Q = Qmax[:bs]\n",
-    "            K = Kmax[:bs]\n",
-    "            def run():\n",
-    "                torch.bmm(Q.view(bs * n, 1, d), K.view(bs * n, seqlen, d).transpose(1, 2))\n",
-    "                torch.cuda.synchronize()\n",
-    "            def clear_cache():\n",
-    "                cache.zero_()\n",
-    "                torch.cuda.synchronize()\n",
-    "            latency = benchmark(run, f_setup=clear_cache, min_repeat=20, min_secs=2)\n",
-    "            l = tail_mean(latency)\n",
-    "            out.append({\n",
-    "                \"n\": n,\n",
-    "                \"d\": d,\n",
-    "                \"seqlen\": seqlen,\n",
-    "                \"bs\": bs,\n",
-    "                \"latency\": l\n",
-    "            })\n",
-    "            pbar.update()\n",
-    "        del cache, Q, K, Qmax, Kmax\n",
-    "        torch.cuda.empty_cache()\n",
-    "    pbar.close()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 10,
-   "id": "3aaad98a",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "data = {}"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 11,
-   "id": "18137de3",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "  0%|          | 0/22 [00:00<?, ?it/s]"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|██████████| 22/22 [00:44<00:00,  2.04s/it, bs=2, d=256, h=3072, n=12, seqlen=256]   \n"
-     ]
-    }
-   ],
-   "source": [
-    "db = []\n",
-    "benchmark_qk_init(db, nd_list, seqlen_list, bs_list)\n",
-    "data[\"qk_init\"] = db"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 12,
-   "id": "26c76e15",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|██████████| 22/22 [00:44<00:00,  2.01s/it, bs=2, d=256, h=3072, n=12, seqlen=256]   \n"
-     ]
-    }
-   ],
-   "source": [
-    "db = []\n",
-    "benchmark_qk_ar(db, nd_list, seqlen_list, bs_list)\n",
-    "data[\"qk_ar\"] = db"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 13,
-   "id": "313e36eb",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "  0%|          | 0/44 [00:00<?, ?it/s, bs=2048, d=256, h=768, n=3, seqlen=256]"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "2048 3 256 256\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      " 25%|██▌       | 11/44 [00:22<01:06,  2.00s/it, bs=2048, d=256, h=768, n=3, seqlen=1] "
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "2048 3 256 1\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      " 50%|█████     | 22/44 [00:44<00:44,  2.00s/it, bs=2048, d=256, h=3072, n=12, seqlen=256]"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "2048 12 256 256\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      " 75%|███████▌  | 33/44 [01:07<00:22,  2.02s/it, bs=2048, d=256, h=3072, n=12, seqlen=1]  "
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "2048 12 256 1\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|██████████| 44/44 [01:29<00:00,  2.03s/it, bs=2, d=256, h=3072, n=12, seqlen=1]   \n"
-     ]
-    }
-   ],
-   "source": [
-    "db = []\n",
-    "benchmark_dense(db, nd_list, seqlen_list, bs_list)\n",
-    "data[\"dense\"] = db"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 14,
-   "id": "50c37959",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "with gzip.open(\"data/20230516-transformer-batching1.pkl.gz\", \"wb\") as f:\n",
-    "    pickle.dump(data, f)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 15,
-   "id": "828ddb54",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "df_dense = (\n",
-    "    pd.DataFrame.from_dict(data[\"dense\"])\n",
-    "    .assign(h=lambda x: x[\"n\"] * x[\"d\"])\n",
-    "    .assign(flop=lambda x: (x[\"bs\"] * x[\"seqlen\"] * x[\"h\"]**2) * 2)\n",
-    "    .assign(io=lambda x: (x[\"bs\"]*x[\"seqlen\"]*x[\"h\"]*2 + x[\"h\"]**2) * 2/x['latency']/1e9)\n",
-    "    .assign(intensity=lambda x: x[\"flop\"] / x[\"io\"])\n",
-    "    .assign(throughput=lambda x: x[\"flop\"] / x[\"latency\"])\n",
-    "    .assign(series=\"dense\")\n",
-    ")\n",
-    "df_qk_init = (\n",
-    "    pd.DataFrame.from_dict(data[\"qk_init\"])\n",
-    "    .assign(h=lambda x: x[\"n\"] * x[\"d\"])\n",
-    "    .assign(flop=lambda x: (x[\"bs\"]*x[\"n\"]*x[\"d\"]*x[\"seqlen\"]**2) * 2)\n",
-    "    .assign(io=lambda x: (x[\"bs\"]*x[\"n\"]*(x[\"seqlen\"]*x[\"d\"]*2 + x[\"seqlen\"]**2)) * 2/x['latency']/1e9)\n",
-    "    .assign(intensity=lambda x: x[\"flop\"] / x[\"io\"])\n",
-    "    .assign(throughput=lambda x: x[\"flop\"] / x[\"latency\"])\n",
-    "    .assign(series=\"qk_init\")\n",
-    ")\n",
-    "df_qk_ar = (\n",
-    "    pd.DataFrame.from_dict(data[\"qk_ar\"])\n",
-    "    .assign(h=lambda x: x[\"n\"] * x[\"d\"])\n",
-    "    .assign(flop=lambda x: (x[\"bs\"]*x[\"n\"]*x[\"d\"]*x[\"seqlen\"]) * 2)\n",
-    "    .assign(io=lambda x: (x[\"bs\"]*x[\"n\"]*(x[\"d\"] + x[\"seqlen\"]*x[\"d\"] + x[\"seqlen\"])) * 2)\n",
-    "    .assign(intensity=lambda x: x[\"flop\"] / x[\"io\"])\n",
-    "    .assign(throughput=lambda x: x[\"bs\"] / x[\"latency\"])\n",
-    "    .assign(series=\"qk_ar\")\n",
-    ")\n",
-    "pd.concat([df_dense, df_qk_init, df_qk_ar]).to_csv(\"data/transformer-batching-microbenchmarks.csv\", index=False)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 39,
-   "id": "c296a395",
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "<module 'pandas' from '/home/ubuntu/Power-RAG/.venv/lib/python3.10/site-packages/pandas/__init__.py'>"
-      ]
-     },
-     "execution_count": 39,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "pd\n"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "a25cdd5a",
-   "metadata": {},
-   "outputs": [],
-   "source": []
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "63b8a531",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import transformers"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "af90eff1",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def _gen_opt_cfg(n_layers: int, d_model: int, n_heads: int, **kwargs) -> transformers.OPTConfig:\n",
-    "    return transformers.OPTConfig(\n",
-    "        num_hidden_layers=n_layers,\n",
-    "        hidden_size=d_model,\n",
-    "        ffn_dim=d_model*4,\n",
-    "        num_attention_heads=n_heads,\n",
-    "        **kwargs\n",
-    "    )\n",
-    "optcfg = {\n",
-    "    # https://arxiv.org/pdf/2205.01068.pdf   Table 2.1\n",
-    "    \"125m\": _gen_opt_cfg(12,   768, 12),\n",
-    "    \"350m\": _gen_opt_cfg(24,  1024, 16),\n",
-    "    \"760m\": _gen_opt_cfg(24,  1536, 16),\n",
-    "    \"1.3b\": _gen_opt_cfg(24,  2048, 32),\n",
-    "    \"2.7b\": _gen_opt_cfg(32,  2560, 32),\n",
-    "    \"6.7b\": _gen_opt_cfg(32,  4096, 32),\n",
-    "     \"13b\": _gen_opt_cfg(40,  5120, 40),\n",
-    "     \"13b_1layer\": _gen_opt_cfg(1,  5120, 40),\n",
-    "     \"30b\": _gen_opt_cfg(48,  7168, 56),\n",
-    "     \"66b\": _gen_opt_cfg(64,  9216, 72),\n",
-    "    \"175b\": _gen_opt_cfg(96, 12288, 96),\n",
-    "    \"175b_1layer\": _gen_opt_cfg(1, 12288, 96),\n",
-    "}"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "5b9ebbec",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def greedy_sample_one(model, input_ids, attention_mask=None, past_key_values=None):\n",
-    "    bs, tgt_len = input_ids.shape\n",
-    "    if past_key_values is not None:\n",
-    "        _bs, _num_heads, src_len, _head_dims = past_key_values[0][0].shape\n",
-    "        assert bs == _bs\n",
-    "    else:\n",
-    "        src_len = 0\n",
-    "    if attention_mask is None:\n",
-    "        attention_mask = torch.ones((bs, src_len + tgt_len), device=model.device)\n",
-    "    ret = model(\n",
-    "        input_ids=input_ids,\n",
-    "        attention_mask=attention_mask,\n",
-    "        past_key_values=past_key_values,\n",
-    "        use_cache=True, output_hidden_states=False, return_dict=True,\n",
-    "    )\n",
-    "    return ret\n",
-    "\n",
-    "def time_greedy_generate(model, input_ids, new_tokens):\n",
-    "    ts = []\n",
-    "    output = input_ids\n",
-    "    past_key_values = None\n",
-    "    cache = torch.empty(int(256e6 // 4), dtype=torch.int, device=model.device)\n",
-    "    attention_mask = torch.ones(input_ids.shape, device=model.device)    \n",
-    "    for _ in range(new_tokens):\n",
-    "        cache.zero_()\n",
-    "        torch.cuda.synchronize()\n",
-    "        st = time.perf_counter_ns()\n",
-    "        \n",
-    "        ret = greedy_sample_one(model, input_ids, attention_mask, past_key_values)\n",
-    "        input_ids = torch.argmax(ret.logits[:, -1, :], axis=-1)[:, None]\n",
-    "        output = torch.cat([output, input_ids], axis=1)\n",
-    "        past_key_values = ret.past_key_values\n",
-    "        attention_mask = torch.cat([attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1)\n",
-    "        \n",
-    "        torch.cuda.synchronize()\n",
-    "        ed = time.perf_counter_ns()\n",
-    "        ts.append((ed-st)/1e9)\n",
-    "    return np.array(ts)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "fc92f940",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "opt_config = optcfg[\"6.7b\"]\n",
-    "\n",
-    "torch.set_default_dtype(torch.bfloat16)\n",
-    "with transformers.modeling_utils.no_init_weights():\n",
-    "    model = transformers.models.opt.OPTForCausalLM(opt_config).to(\"cuda\")\n",
-    "torch.set_default_dtype(torch.float32)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "c19fa396",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "db = {}\n",
-    "input_tokens = 200\n",
-    "new_tokens = 500\n",
-    "for bs in tqdm(list(itertools.chain(range(1, 8), range(8, 16, 2), [16]))):\n",
-    "    x = torch.randint(1000, 10000, (bs, input_tokens), device=model.device)\n",
-    "    stack = []\n",
-    "    for _ in range(10):\n",
-    "        l = time_greedy_generate(model, x, new_tokens=new_tokens)\n",
-    "        stack.append(l)\n",
-    "    db[bs] = np.median(np.stack(stack), axis=0)\n",
-    "    del x\n",
-    "    torch.cuda.empty_cache()\n",
-    "del model\n",
-    "torch.cuda.empty_cache()\n",
-    "\n",
-    "with gzip.open(\"data/20230516-e2e-text-generation-batch.pkl.gz\", \"wb\") as f:\n",
-    "    pickle.dump(db, f)"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": ".venv",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.10.12"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 5
-}

research/paper_plot/acc_fig.py

@@ -1,165 +0,0 @@
-import matplotlib.pyplot as plt
-import numpy as np
-import pandas as pd
-
-# Set plot parameters
-plt.rcParams["font.family"] = "Helvetica"
-plt.rcParams["ytick.direction"] = "in"
-plt.rcParams["hatch.linewidth"] = 1.5
-plt.rcParams["font.weight"] = "bold"
-plt.rcParams["axes.labelweight"] = "bold"
-plt.rcParams["text.usetex"] = True
-
-# Path settings
-FIGURE_PATH = "./paper_plot/figures"
-
-# Load accuracy data
-acc_data = pd.read_csv("./paper_plot/data/acc.csv")
-
-# Create figure with 4 subplots (one for each dataset)
-fig, axs = plt.subplots(1, 4)
-fig.set_size_inches(9, 2.5)
-
-# Reduce the spacing between subplots
-# plt.subplots_adjust(wspace=0.2)  # Reduced from 0.3 to 0.1
-
-# Define datasets and their columns
-datasets = ["NQ", "TriviaQA", "GPQA", "HotpotQA"]
-metrics = ["Exact Match", "F1"]
-
-# Define bar settings - make bars thicker
-# total_width, n = 0.9, 3  # increased total width and n for three models
-# width = total_width / n
-# The 'width' variable below now defines the distance between the centers of adjacent bars within a group.
-# It's also used as the base for calculating the actual plotted bar width.
-# Original 2 bars had centers 1.0 apart. For 3 bars, we need a smaller distance.
-# A value of 0.64 for distance between centers, with a scaling factor of 0.8 for bar width,
-# results in an actual bar width of ~0.51, and a group span of ~1.79, similar to original's ~1.76.
-n = 3 # Number of models
-width = 0.64  # Distance between centers of adjacent bars in a group
-bar_width_plotting_factor = 0.8 # Bar takes 80% of the space defined by 'width'
-
-# Colors and hatches
-edgecolors = ["dimgrey", "#63B8B6", "tomato"]  # Added color for PQ 5
-hatches = ["/////", "xxxxx", "\\\\\\\\\\"]  # Added hatch for PQ 5
-labels = ["BM25", "PQ Compressed", "Ours"] # Added PQ 5
-
-# Create plots for each dataset
-for i, dataset in enumerate(datasets):
-    ax = axs[i]
-    
-    # Get data for this dataset and convert to percentages
-    em_values = [
-        acc_data.loc[0, f"{dataset} Exact Match"] * 100, 
-        acc_data.loc[1, f"{dataset} Exact Match"] * 100,
-        acc_data.loc[2, f"{dataset} Exact Match"] * 100  # Added PQ 5 EM data
-    ]
-    f1_values = [
-        acc_data.loc[0, f"{dataset} F1"] * 100, 
-        acc_data.loc[1, f"{dataset} F1"] * 100,
-        acc_data.loc[2, f"{dataset} F1"] * 100  # Added PQ 5 F1 data
-    ]
-    
-    # Define x positions for bars
-    # For EM: center - width, center, center + width
-    # For F1: center - width, center, center + width
-    group_centers = [1.0, 3.0] # Centers for EM and F1 groups
-    bar_offsets = [-width, 0, width]
-
-    # Plot all bars on the same axis
-    for metric_idx, metric_group_center in enumerate(group_centers):
-        values_to_plot = em_values if metric_idx == 0 else f1_values
-        for j, model_label in enumerate(labels):
-            x_pos = metric_group_center + bar_offsets[j]
-            bar_value = values_to_plot[j]
-            
-            ax.bar(
-                x_pos,
-                bar_value,
-                width=width * bar_width_plotting_factor, # Use the new factor for bar width
-                color="white",
-                edgecolor=edgecolors[j],
-                hatch=hatches[j],
-                linewidth=1.5,
-                label=model_label if i == 0 and metric_idx == 0 else None # Label only once
-            )
-            
-            # Add value on top of bar
-            ax.text(x_pos, bar_value + (0.1 if dataset == "GPQA" else 0.1), 
-                    f"{bar_value:.1f}", ha='center', va='bottom', 
-                    fontsize=9, fontweight='bold') # Reduced fontsize for text on bars
-    
-    # Set x-ticks and labels
-    ax.set_xticks(group_centers)  # Position ticks at the center of each group
-    xticklabels = ax.set_xticklabels(metrics, fontsize=12)
-
-    # Now, shift these labels slightly to the right
-    # Adjust this value to control the amount of shift (in data coordinates)
-    # Given your group_centers are 1.0 and 3.0, a small value like 0.05 to 0.15 might be appropriate.
-    # horizontal_shift = 0.7  # Try adjusting this value
-
-    # for label in xticklabels:
-    #     # Get the current x position (which is the tick location)
-    #     current_x_pos = label.get_position()[0]
-    #     # Set the new x position by adding the shift
-    #     label.set_position((current_x_pos + horizontal_shift, label.get_position()[1]))
-    #     # Ensure the label remains horizontally centered on this new x position
-    #     # (set_xticklabels defaults to 'center', so this re-affirms it if needed)
-    #     label.set_horizontalalignment('center')
-
-    # Set title
-    ax.set_title(dataset, fontsize=14)
-    
-    # Set y-label for all subplots
-    if i == 0:
-        ax.set_ylabel("Accuracy (\%)", fontsize=12, fontweight="bold")
-    else:
-        # Hide y-tick labels for non-first subplots to save space
-        ax.tick_params(axis='y', labelsize=10)
-    
-    # Set y-limits based on data range
-    all_values = em_values + f1_values
-    max_val = max(all_values)
-    min_val = min(all_values)
-    
-    # Special handling for GPQA which has very low values
-    if dataset == "GPQA":
-        ax.set_ylim(0, 10.0)  # Set a fixed range for GPQA
-    else:
-        # Reduce the extra space above the bars
-        ax.set_ylim(min_val * 0.9, max_val * 1.1) # Adjusted upper limit for text
-    
-    # Format y-ticks as percentages
-    ax.yaxis.set_major_formatter(plt.FuncFormatter(lambda y, _: ' {:.0f}'.format(y)))
-    
-    # Set x-limits to properly space the bars with less blank space
-    # ax.set_xlim(group_centers[0] - total_width, group_centers[1] + total_width)
-    # Set xlim to be similar to original (0,4) for group_centers (1,3) => margin of 1.0
-    ax.set_xlim(group_centers[0] - 1.0, group_centers[1] + 1.0)
-    
-    # Add a box around the subplot
-    # for spine in ax.spines.values():
-    #     spine.set_visible(True)
-    #     spine.set_linewidth(1.0)
-    
-    # Add legend to first subplot
-    if i == 0:
-        ax.legend(
-            bbox_to_anchor=(2.21, 1.35), # Adjusted anchor if needed
-            ncol=3, # Changed to 3 columns for three labels
-            loc="upper center",
-            labelspacing=0.1,
-            edgecolor="black",
-            facecolor="white",
-            framealpha=1,
-            shadow=False,
-            fancybox=False,
-            handlelength=1.0,
-            handletextpad=0.6,
-            columnspacing=0.8,
-            prop={"weight": "bold", "size": 12},
-        )
-
-# Save figure with tight layout but no additional padding
-plt.savefig(FIGURE_PATH + "/accuracy_comparison.pdf", bbox_inches='tight', pad_inches=0.05)
-plt.show()

research/paper_plot/analyze_visits.py

@@ -1,309 +0,0 @@
-#!/usr/bin/env python3
-# -*- coding:utf-8 -*-
-# Motto: Were It to Benefit My Country, I Would Lay Down My Life!
-# \file: /hnsw_degree_visit_plot_binned_academic.py
-# \brief: Generates a binned bar plot of HNSW node average per-query visit probability
-#         per degree bin, styled for academic publications, with caching.
-# Author: raphael hao (Original script by user, styling and caching adapted by Gemini)
-
-# %%
-import numpy as np
-import pandas as pd
-import matplotlib.pyplot as plt
-import re
-from collections import Counter
-import os # For robust filepath manipulation
-import math # For calculating scaling factor
-import pickle # For caching data
-
-# %%
-# --- Matplotlib parameters for academic paper style (from reference) ---
-plt.rcParams["font.family"] = "Helvetica"
-plt.rcParams["ytick.direction"] = "in"
-plt.rcParams["hatch.linewidth"] = 1.5
-plt.rcParams["font.weight"] = "bold"
-plt.rcParams["axes.labelweight"] = "bold"
-plt.rcParams["text.usetex"] = True # Use LaTeX for text rendering (if available)
-
-# --- Define styles from reference ---
-edgecolors_ref = ["dimgrey", "#63B8B6", "tomato", "silver", "slategray"]
-
-# %%
-# --- File Paths ---
-degree_file = '/opt/dlami/nvme/scaling_out/indices/rpj_wiki/facebook/contriever-msmarco/hnsw/degree_distribution.txt'
-visit_log_file = './re.log'
-output_image_file = './paper_plot/figures/hnsw_visit_count_per_degree_corrected.pdf'
-# --- CACHE FILE PATH: Keep this consistent ---
-CACHE_FILE_PATH = './binned_plot_data_cache.pkl'
-
-# --- Configuration ---
-# Set to True to bypass cache and force recomputation.
-# Otherwise, delete CACHE_FILE_PATH manually to force recomputation.
-FORCE_RECOMPUTE = False
-NUMBER_OF_QUERIES = 1000.0 # Number of queries the visit_counts are based on
-
-# Create directory for figures if it doesn't exist
-output_dir = os.path.dirname(output_image_file)
-if output_dir and not os.path.exists(output_dir):
-    os.makedirs(output_dir)
-    print(f"Created directory: {output_dir}")
-
-# %%
-# --- Attempt to load data from cache or compute ---
-df_plot_data = None
-bin_size_for_plot = None # Will hold the bin_size associated with df_plot_data
-
-if not FORCE_RECOMPUTE and os.path.exists(CACHE_FILE_PATH):
-    try:
-        with open(CACHE_FILE_PATH, 'rb') as f:
-            cache_content = pickle.load(f)
-            df_plot_data = cache_content['data']
-            bin_size_for_plot = cache_content['bin_size']
-        # Basic validation of cached data
-        # Expecting 'average_visit_count_per_node_in_bin' (raw average over NUMBER_OF_QUERIES)
-        if not isinstance(df_plot_data, pd.DataFrame) or \
-           'degree_bin_label' not in df_plot_data.columns or \
-           'average_visit_count_per_node_in_bin' not in df_plot_data.columns or \
-           not isinstance(bin_size_for_plot, int):
-            print("Cached data is not in the expected format or missing 'average_visit_count_per_node_in_bin'. Recomputing.")
-            df_plot_data = None # Invalidate to trigger recomputation
-        else:
-            print(f"Successfully loaded binned data from cache: {CACHE_FILE_PATH}")
-
-        # --- Modify the label loaded from cache for display purpose ---
-        # This modification only happens when data is loaded from cache and meets specific conditions.
-        # Assumption: If the bin_size_for_plot in cache is 5,
-        # then the original label "0-4" actually represents nodes with degree 1-4 (because you guarantee no 0-degree nodes).
-        if df_plot_data is not None and 'degree_bin_label' in df_plot_data.columns and bin_size_for_plot == 5:
-            # Check if "0-4" label exists
-            if '0-4' in df_plot_data['degree_bin_label'].values:
-                # Use .loc to ensure the modification is on the original DataFrame
-                df_plot_data.loc[df_plot_data['degree_bin_label'] == '0-4', 'degree_bin_label'] = '1-4'
-                print("Modified degree_bin_label from '0-4' to '1-4' for display purpose.")
-    except Exception as e:
-        print(f"Error loading from cache: {e}. Recomputing.")
-        df_plot_data = None # Invalidate to trigger recomputation
-
-if df_plot_data is None:
-    print("Cache not found, invalid, or recompute forced. Computing data from scratch...")
-    # --- 1. Read Degree Distribution File ---
-    degrees_data = []
-    try:
-        with open(degree_file, 'r') as f:
-            for i, line in enumerate(f):
-                line_stripped = line.strip()
-                if line_stripped:
-                    degrees_data.append({'node_id': i, 'degree': int(line_stripped)})
-    except FileNotFoundError:
-        print(f"Error: Degree file '{degree_file}' not found. Using dummy data for degrees.")
-        degrees_data = [{'node_id': i, 'degree': (i % 20) + 1 } for i in range(200)]
-        degrees_data.extend([{'node_id': 200+i, 'degree': i} for i in range(58, 67)]) # For 60-64 bin
-        degrees_data.extend([{'node_id': 300+i, 'degree': (i % 5)+1} for i in range(10)]) # Low degrees
-        degrees_data.extend([{'node_id': 400+i, 'degree': 80 + (i%5)} for i in range(10)]) # High degrees
-
-
-    if not degrees_data:
-        print(f"Critical Error: No data loaded or generated for degrees. Exiting.")
-        exit()
-    df_degrees = pd.DataFrame(degrees_data)
-    print(f"Successfully loaded/generated {len(df_degrees)} degree entries.")
-
-    # --- 2. Read Visit Log File and Count Frequencies ---
-    visit_counts = Counter()
-    node_id_pattern = re.compile(r"Vis(i)?ted node: (\d+)")
-    try:
-        with open(visit_log_file, 'r') as f_log:
-            for line_num, line in enumerate(f_log, 1):
-                match = node_id_pattern.search(line)
-                if match:
-                    try:
-                        node_id = int(match.group(2))
-                        visit_counts[node_id] += 1 # Increment visit count for the node
-                    except ValueError:
-                        print(f"Warning: Non-integer node_id in log '{visit_log_file}' line {line_num}: {line.strip()}")
-    except FileNotFoundError:
-        print(f"Warning: Visit log file '{visit_log_file}' not found. Using dummy visit counts.")
-        if not df_degrees.empty:
-            for node_id_val in df_degrees['node_id'].sample(frac=0.9, random_state=1234): # Seed for reproducibility
-                degree_val = df_degrees[df_degrees['node_id'] == node_id_val]['degree'].iloc[0]
-                # Generate visit counts to test different probability magnitudes
-                if node_id_val % 23 == 0: # Very low probability
-                     lambda_val = 0.0005 * (100 / (max(1,degree_val) + 1)) # avg visits over 1k queries
-                elif node_id_val % 11 == 0: # Low probability
-                     lambda_val = 0.05 * (100 / (max(1,degree_val) + 1))
-                elif node_id_val % 5 == 0: # Moderate probability
-                     lambda_val = 2.5 * (100 / (max(1,degree_val) + 1))
-                else: # Higher probability (but still < 1000 visits for a single node usually)
-                     lambda_val = 50 * (100 / (max(1,degree_val) + 1))
-                visit_counts[node_id_val] = np.random.poisson(lambda_val)
-                if visit_counts[node_id_val] < 0: visit_counts[node_id_val] = 0
-
-    if not visit_counts:
-        print(f"Warning: No visit data parsed/generated. Plot may show zero visits.")
-        df_visits = pd.DataFrame(columns=['node_id', 'visit_count'])
-    else:
-        df_visits_list = [{'node_id': nid, 'visit_count': count} for nid, count in visit_counts.items()]
-        df_visits = pd.DataFrame(df_visits_list)
-    print(f"Parsed/generated {len(df_visits)} unique visited nodes, totaling {sum(visit_counts.values())} visits (simulated over {NUMBER_OF_QUERIES} queries).")
-
-    # --- 3. Merge Degree Data with Visit Data ---
-    df_merged = pd.merge(df_degrees, df_visits, on='node_id', how='left')
-    df_merged['visit_count'] = df_merged['visit_count'].fillna(0).astype(float) # visit_count is total over NUMBER_OF_QUERIES
-    print(f"Merged data contains {len(df_merged)} entries.")
-
-    # --- 5. Binning Degrees and Calculating Average Visit Count per Node in Bin (over NUMBER_OF_QUERIES) ---
-    current_bin_size = 5
-    bin_size_for_plot = current_bin_size
-
-    if not df_degrees.empty:
-        print(f"\nBinning degrees into groups of {current_bin_size} for average visit count calculation...")
-
-        df_merged_with_bins = df_merged.copy()
-        df_merged_with_bins['degree_bin_start'] = (df_merged_with_bins['degree'] // current_bin_size) * current_bin_size
-        
-        df_binned_analysis = df_merged_with_bins.groupby('degree_bin_start').agg(
-            total_visit_count_in_bin=('visit_count', 'sum'),
-            node_count_in_bin=('node_id', 'nunique')
-        ).reset_index()
-
-        # This is the average number of times a node in this bin was visited over NUMBER_OF_QUERIES queries.
-        # This value is what gets cached.
-        df_binned_analysis['average_visit_count_per_node_in_bin'] = 0.0
-        df_binned_analysis.loc[df_binned_analysis['node_count_in_bin'] > 0, 'average_visit_count_per_node_in_bin'] = \
-            df_binned_analysis['total_visit_count_in_bin'] / df_binned_analysis['node_count_in_bin']
-        
-        df_binned_analysis['degree_bin_label'] = df_binned_analysis['degree_bin_start'].astype(str) + '-' + \
-                                                 (df_binned_analysis['degree_bin_start'] + current_bin_size - 1).astype(str)
-        
-        bin_to_drop_label = '60-64'
-        original_length = len(df_binned_analysis)
-        df_plot_data_intermediate = df_binned_analysis[df_binned_analysis['degree_bin_label'] != bin_to_drop_label].copy()
-        if len(df_plot_data_intermediate) < original_length:
-            print(f"\nManually dropped the bin: '{bin_to_drop_label}'")
-        else:
-            print(f"\nNote: Bin '{bin_to_drop_label}' not found for dropping or already removed.")
-        
-        df_plot_data = df_plot_data_intermediate
-        
-        print(f"\nBinned data (average visit count per node in bin over {NUMBER_OF_QUERIES} queries) for plotting prepared:")
-        print(df_plot_data[['degree_bin_label', 'average_visit_count_per_node_in_bin']].head())
-
-        if df_plot_data is not None and not df_plot_data.empty:
-            try:
-                with open(CACHE_FILE_PATH, 'wb') as f:
-                    pickle.dump({'data': df_plot_data, 'bin_size': bin_size_for_plot}, f)
-                print(f"Saved computed binned data to cache: {CACHE_FILE_PATH}")
-            except Exception as e:
-                print(f"Error saving data to cache: {e}")
-        elif df_plot_data is None or df_plot_data.empty:
-             print("Computed data for binned plot is empty, not saving to cache.")
-    else:
-        print("Degree data (df_degrees) is empty. Cannot perform binning.")
-        df_plot_data = pd.DataFrame()
-        bin_size_for_plot = current_bin_size
-
-# %%
-# --- 6. Plotting (Binned Bar Chart - Academic Style) ---
-
-if df_plot_data is not None and not df_plot_data.empty and 'average_visit_count_per_node_in_bin' in df_plot_data.columns:
-    base_name, ext = os.path.splitext(output_image_file)
-    # --- OUTPUT PDF FILE NAME: Keep this consistent ---
-    binned_output_image_file = base_name + ext
-
-    fig, ax = plt.subplots(figsize=(6, 2.5)) # Adjusted figure size
-
-    df_plot_data_plotting = df_plot_data.copy()
-    # Calculate per-query probability: (avg visits over N queries) / N
-    df_plot_data_plotting['per_query_visit_probability'] = \
-        df_plot_data_plotting['average_visit_count_per_node_in_bin'] / NUMBER_OF_QUERIES
-    
-    max_probability = df_plot_data_plotting['per_query_visit_probability'].max()
-    
-    y_axis_values_to_plot = df_plot_data_plotting['per_query_visit_probability']
-    y_axis_label = r"Per-Query Node Visit Probability in Bin" # Base label
-
-    apply_scaling_to_label_and_values = False # Initialize flag
-    exponent_for_label_display = 0 # Initialize exponent
-
-    if pd.notna(max_probability) and max_probability > 0:
-        potential_exponent = math.floor(math.log10(max_probability))
-        
-        if potential_exponent <= -4 or potential_exponent >= 0: 
-            apply_scaling_to_label_and_values = True
-            exponent_for_label_display = potential_exponent
-            # No specific adjustment for potential_exponent >=0 here, it's handled by the general logic.
-
-        if apply_scaling_to_label_and_values:
-            y_axis_label = rf"Visit Probability ($\times 10^{{{exponent_for_label_display}}}$)"
-            y_axis_values_to_plot = df_plot_data_plotting['per_query_visit_probability'] / (10**exponent_for_label_display)
-            print(f"Plotting with Max per-query probability: {max_probability:.2e}, Exponent for label: {exponent_for_label_display}. Y-axis values scaled for plot.")
-        else:
-            print(f"Plotting with Max per-query probability: {max_probability:.2e}. Plotting direct probabilities without label scaling (exponent {potential_exponent} is within no-scale range [-3, -1]).")
-
-    elif pd.notna(max_probability) and max_probability == 0:
-        print("Max per-query probability is 0. Plotting direct probabilities (all zeros).")
-    else:
-        print(f"Max per-query probability is NaN or invalid ({max_probability}). Plotting direct probabilities without scaling if possible.")
-    
-    ax.bar(
-        df_plot_data_plotting['degree_bin_label'],
-        y_axis_values_to_plot,
-        color='white',
-        edgecolor=edgecolors_ref[0],
-        linewidth=1.5,
-        width=0.8
-    )
-    
-    ax.set_xlabel('Node Degree', fontsize=10.5, labelpad=6)
-    # MODIFIED LINE: Added labelpad to move the y-axis label to the left
-    ax.set_ylabel(y_axis_label, fontsize=10.5, labelpad=10) 
-
-    ax.yaxis.set_major_formatter(plt.FuncFormatter(lambda x, pos: f"{x:.0f}%"))
-
-    num_bins = len(df_plot_data_plotting)
-    if num_bins > 12: 
-        ax.set_xticks(ax.get_xticks())
-        ax.set_xticklabels(ax.get_xticklabels(), rotation=45, ha="right", fontsize=9)
-    elif num_bins > 8:
-         ax.tick_params(axis='x', labelsize=9)
-    else:
-        ax.tick_params(axis='x', labelsize=10)
-    
-    ax.tick_params(axis='y', labelsize=10)
-
-    padding_factor = 0.05
-    current_max_y_on_axis = y_axis_values_to_plot.max()
-    
-    upper_y_limit = 0.1 # Default small upper limit
-    if pd.notna(current_max_y_on_axis):
-        if current_max_y_on_axis > 0:
-             # Adjust minimum visible range based on whether scaling was applied and the exponent
-            min_meaningful_limit = 0.01
-            if apply_scaling_to_label_and_values and exponent_for_label_display >= 0 : # Numbers on axis are smaller due to positive exponent scaling
-                 min_meaningful_limit = 0.1 # If original numbers were e.g. 2500 (2.5 x 10^3), scaled axis is 2.5, 0.1 is fine
-            elif not apply_scaling_to_label_and_values and pd.notna(max_probability) and max_probability >=1: # Direct large probabilities
-                 min_meaningful_limit = 1 # If max prob is 2.5 (250%), axis value 2.5, needs larger base limit
-            
-            upper_y_limit = max(min_meaningful_limit, current_max_y_on_axis * (1 + padding_factor))
-
-        else: # current_max_y_on_axis is 0
-            upper_y_limit = 0.1 
-        ax.set_ylim(0, upper_y_limit)
-    else:
-        ax.set_ylim(0, 1.0) # Default for empty or NaN data
-
-    plt.tight_layout()
-    plt.savefig(binned_output_image_file, bbox_inches="tight", dpi=300)
-    print(f"Binned bar chart saved to {binned_output_image_file}")
-    plt.show()
-    plt.close(fig)
-else:
-    if df_plot_data is None:
-        print("Data for plotting (df_plot_data) is None. Skipping plot generation.")
-    elif df_plot_data.empty:
-        print("Data for plotting (df_plot_data) is empty. Skipping plot generation.")
-    elif 'average_visit_count_per_node_in_bin' not in df_plot_data.columns: 
-        print("Essential column 'average_visit_count_per_node_in_bin' is missing in df_plot_data. Skipping plot generation.")
-
-# %%
-print("Script finished.")

research/paper_plot/b.md

@@ -1,7 +0,0 @@
-In this paper, we present LiteANN, a storage-efficient approximate nearest neighbor (ANN) search index optimized for resource-constrained personal devices. LiteANN combines a compact graph-based structure with an efficient on-the-fly recomputation strategy to enable fast and accurate retrieval wih minimal storage overhead. Our evaluation shows that LiteANN reduces index size to under 5% of the original raw data – up to 50× smaller than standard indexes – while achieving 90% top-3 recall in under 2 seconds on real-world question-answering benchmarks.
-
-
-
-
-
-

research/paper_plot/cache_degree_data.py

@@ -1,81 +0,0 @@
-import numpy as np
-import os
-
-# --- Configuration for Data Paths and Labels (Mirrors plotting script for consistency) ---
-BIG_GRAPH_PATHS = [
-    "/opt/dlami/nvme/scaling_out/indices/rpj_wiki/facebook/contriever-msmarco/hnsw/",
-    "/opt/dlami/nvme/scaling_out/embeddings/facebook/contriever-msmarco/rpj_wiki/1-shards/indices/99_4_degree_based_hnsw_IP_M32_efC256/",
-    "/opt/dlami/nvme/scaling_out/embeddings/facebook/contriever-msmarco/rpj_wiki/1-shards/indices/d9_hnsw_IP_M8_efC128/",
-    "/opt/dlami/nvme/scaling_out/embeddings/facebook/contriever-msmarco/rpj_wiki/1-shards/indices/half_edges_IP_M32_efC128/"
-]
-STATS_FILE_NAME = "degree_distribution.txt"
-BIG_GRAPH_LABELS = [  # These will be used as keys in the cached file
-    "HNSW-Base",
-    "DegreeGuide",
-    "HNSW-D9", 
-    "RandCut",
-]
-# Average degrees are static and can be directly used in the plotting script or also cached.
-# For simplicity here, we'll focus on caching the dynamic degree arrays.
-# BIG_GRAPH_AVG_DEG = [18, 9, 9, 9] 
-
-# --- Cache File Configuration ---
-DATA_CACHE_DIR = "./paper_plot/data/"
-CACHE_FILE_NAME = "big_graph_degree_data.npz" # Using .npz for multiple arrays
-
-def create_degree_data_cache():
-    """
-    Reads degree distribution data from specified text files and saves it
-    into a compressed NumPy (.npz) cache file.
-    """
-    os.makedirs(DATA_CACHE_DIR, exist_ok=True)
-    cache_file_path = os.path.join(DATA_CACHE_DIR, CACHE_FILE_NAME)
-
-    cached_data = {}
-    print(f"Starting data caching process for {len(BIG_GRAPH_PATHS)} graph types...")
-
-    for i, base_path in enumerate(BIG_GRAPH_PATHS):
-        method_label = BIG_GRAPH_LABELS[i]
-        degree_file_path = os.path.join(base_path, STATS_FILE_NAME)
-        
-        print(f"Processing: {method_label} from {degree_file_path}")
-        
-        try:
-            # Load degrees as integers
-            degrees = np.loadtxt(degree_file_path, dtype=int)
-            
-            if degrees.size == 0:
-                print(f"  [WARN] Degree file is empty: {degree_file_path}. Storing as empty array for {method_label}.")
-                # Store an empty array or handle as needed. For npz, an empty array is fine.
-                cached_data[method_label] = np.array([], dtype=int) 
-            else:
-                # Store the loaded degrees array with the method label as the key
-                cached_data[method_label] = degrees
-                print(f"  [INFO] Loaded {len(degrees)} degrees for {method_label}. Max degree: {np.max(degrees) if degrees.size > 0 else 'N/A'}")
-                
-        except FileNotFoundError:
-            print(f"  [ERROR] Degree file not found: {degree_file_path}. Skipping {method_label}.")
-            # Optionally store a placeholder or skip. For robustness, store None or an empty array.
-            # Storing None might require special handling when loading. Empty array is safer for np.load.
-            cached_data[method_label] = np.array([], dtype=int) # Store empty array if file not found
-        except Exception as e:
-            print(f"  [ERROR] An error occurred loading {degree_file_path} for {method_label}: {e}")
-            cached_data[method_label] = np.array([], dtype=int) # Store empty array on other errors
-
-    if not cached_data:
-        print("[ERROR] No data was successfully processed or loaded. Cache file will not be created.")
-        return
-
-    try:
-        # Save all collected degree arrays into a single .npz file.
-        # Using savez_compressed for potentially smaller file size.
-        np.savez_compressed(cache_file_path, **cached_data)
-        print(f"\n[SUCCESS] Degree distribution data successfully cached to: {os.path.abspath(cache_file_path)}")
-        print("Cached arrays (keys):", list(cached_data.keys()))
-    except Exception as e:
-        print(f"\n[ERROR] Failed to save data to cache file {cache_file_path}: {e}")
-
-if __name__ == "__main__":
-    print("--- Degree Distribution Data Caching Script ---")
-    create_degree_data_cache()
-    print("--- Caching script finished. ---")

research/paper_plot/data/acc.csv

@@ -1,4 +0,0 @@
-Model,NQ Exact Match,NQ F1,TriviaQA Exact Match,TriviaQA F1,GPQA Exact Match,GPQA F1,HotpotQA Exact Match,HotpotQA F1
-BM25,0.192,0.277,0.406,0.474,0.020089,0.04524,0.162,0.239
-PQ 5,0.2075,0.291,0.422,0.495,0.0201,0.0445,0.148,0.219
-Ours,0.265,0.361,0.533,0.604,0.02008,0.0452,0.182,0.2729

research/paper_plot/data/big_graph_degree_data.npz

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:1296720e79196bbdf38f051043c1b054667803726a24036c0b6a87cedb204ea5
-size 227482438

research/paper_plot/data/branches.csv

@@ -1,21 +0,0 @@
-2,1,512,1024,0.541,0.326,1.659509202
-2,2,512,1024,0.979,0.621,1.576489533
-2,4,512,1024,1.846,0.977,1.889457523
-2,8,512,1024,3.575,1.943,1.83993824
-2,16,512,1024,7.035,3.733,1.884543263
-2,32,512,1024,15.655,8.517,1.838088529
-2,64,512,1024,32.772,17.43,1.88020654
-4,1,512,1024,2.675,1.38,1.938405797
-4,2,512,1024,5.397,2.339,2.307396323
-4,4,512,1024,10.672,4.944,2.158576052
-4,8,512,1024,21.061,9.266,2.272933305
-4,16,512,1024,46.332,18.334,2.527108105
-4,32,512,1024,99.607,36.156,2.754923111
-4,64,512,1024,186.348,72.356,2.575432583
-8,1,512,1024,7.325,4.087,1.792268167
-8,2,512,1024,14.109,7.491,1.883460152
-8,4,512,1024,28.499,14.013,2.033754371
-8,8,512,1024,65.222,27.453,2.375769497
-8,16,512,1024,146.294,52.55,2.783901047
-8,32,512,1024,277.099,103.61,2.674442621
-8,64,512,1024,512.979,208.36,2.461984066

research/paper_plot/data/latency_ablation.csv

@@ -1,9 +0,0 @@
-Dataset,Metric,Original,original + batch,original + two_level,original + two_level + batch
-NQ,Latency,6.9,5.8,4.2,3.7
-NQ,SpeedUp,1,1.18965517,1.64285714,1.86486486
-TriviaQA,Latency,17.054,14.542,12.046,10.83
-TriviaQA,SpeedUp,1,1.17274103,1.41573967,1.57469990
-GPQA,Latency,9.164,7.639,6.798,5.77
-GPQA,SpeedUp,1,1.19963346,1.34804354,1.58821490
-HotpotQA,Latency,60.279,39.827,50.664,29.868
-HotpotQA,SpeedUp,1,1.51352098,1.18977972,2.01817999

research/paper_plot/data/main_latency.csv

@@ -1,25 +0,0 @@
-Dataset,Hardware,Recall_target,HNSW,IVF,DiskANN,IVF-Disk,IVF-Recompute,Our,BM25,LLM_Gen_Time_1B,LLM_Gen_Time_3B,LLM_Gen_Time_7B
-NQ,A10,85%,0.046,1.656,0.017,2.996,482.53,3.323,0.021,0.085,0.217,0.472
-NQ,A10,90%,0.051,2.552,0.028,3.437,769.04,4.616,0,0.085,0.217,0.472
-NQ,A10,95%,0.055,5.163,0.070,5.602,1436.26,19.494,0,0.085,0.217,0.472
-NQ,MAC,85%,0,0,0.152,2.199,1535.10,7.971,0.033,0.316,0.717,1.468
-NQ,MAC,90%,0,0,0.37,2.936,2446.60,13.843,0,0.316,0.717,1.468
-NQ,MAC,95%,0,0,1.207,4.191,4569.29,44.363,0,0.316,0.717,1.468
-TriviaQA,A10,85%,0.042,1.772,0.032,2.464,560.5,3.752,0.033,0.139,0.156,0.315
-TriviaQA,A10,90%,0.043,3.541,0.057,3.651,997.81,5.777,0,0.139,0.156,0.315
-TriviaQA,A10,95%,0.053,7.168,0.090,5.458,2005.33,20.944,0,0.139,0.156,0.315
-TriviaQA,MAC,85%,0,0,0.481,1.875,1783.14787,8.889,0.036,0.325,0.692,1.415
-TriviaQA,MAC,90%,0,0,0.984,2.639,3174.410301,17.145,0,0.325,0.692,1.415
-TriviaQA,MAC,95%,0,0,1.578,3.884,6379.712245,47.909,0,0.325,0.692,1.415
-GPQA,A10,85%,0.041,0.134,0.024,0.048,40.16,1.897,0.137,0.443,0.396,0.651
-GPQA,A10,90%,0.042,0.174,0.034,0.06,54.71,1.733,0,0.443,0.396,0.651
-GPQA,A10,95%,0.045,0.292,0.051,0.11,97.67,4.033,0,0.443,0.396,0.651
-GPQA,MAC,85%,0,0,0.144,0.087,127.7707505,4.762,0.100,0.37,0.813,1.676
-GPQA,MAC,90%,0,0,0.288,0.108,174.0647409,5.223,0,0.37,0.813,1.676
-GPQA,MAC,95%,0,0,0.497,0.132,310.7380142,9.715,0,0.37,0.813,1.676
-HotpotQA,A10,85%,0.044,2.519,0.054,4.048,724.26,10.358,0.70,0.144,0.196,0.420
-HotpotQA,A10,90%,0.049,3.867,0.109,5.045,1173.67,15.515,0,0.144,0.196,0.420
-HotpotQA,A10,95%,0.07,10.928,0.412,8.659,3079.57,61.757,0,0.144,0.196,0.420
-HotpotQA,MAC,85%,0,0,0.974,2.844,2304.125187,23.636,0.052,0.144,0.196,0.420
-HotpotQA,MAC,90%,0,0,1.913,3.542,3415.736201,44.803,0,0.144,0.196,0.420
-HotpotQA,MAC,95%,0,0,5.783,6.764,9797.244043,140.62,0,0.144,0.196,0.420

research/paper_plot/data/main_latency_small.csv

@@ -1,25 +0,0 @@
-Dataset,Hardware,Recall_target,HNSW,IVF,DiskANN,IVF-Disk,IVF-Recompute,Our,
-NQ,A10,85%,0.046,1.656,0.017,2.996,482.53,4.243,
-NQ,A10,90%,0.051,2.552,0.028,3.437,769.04,8.136,
-NQ,A10,95%,0.055,5.163,0.070,5.602,1436.26,27.275,
-NQ,MAC,85%,0,0,0.152,2.199,1535.10,10.672,
-NQ,MAC,90%,0,0,0.37,2.936,2446.60,19.941,
-NQ,MAC,95%,0,0,1.207,4.191,4569.29,61.383,
-TriviaQA,A10,85%,0.042,1.772,0.032,2.464,560.5,5.612,
-TriviaQA,A10,90%,0.043,3.541,0.057,3.651,997.81,10.737,
-TriviaQA,A10,95%,0.053,7.168,0.090,5.458,2005.33,36.387,
-TriviaQA,MAC,85%,0,0,0.481,1.875,1783.14787,12.825,
-TriviaQA,MAC,90%,0,0,0.984,2.639,3174.410301,24.977,
-TriviaQA,MAC,95%,0,0,1.578,3.884,6379.712245,85.734,
-GPQA,A10,85%,0.041,0.134,0.024,0.048,40.16,2.269,
-GPQA,A10,90%,0.042,0.174,0.034,0.06,54.71,3.200,
-GPQA,A10,95%,0.045,0.292,0.051,0.11,97.67,7.445,
-GPQA,MAC,85%,0,0,0.144,0.087,127.7707505,6.123,
-GPQA,MAC,90%,0,0,0.288,0.108,174.0647409,8.507,
-GPQA,MAC,95%,0,0,0.497,0.132,310.7380142,19.577,
-HotpotQA,A10,85%,0.044,2.519,0.054,4.048,724.26,14.713,
-HotpotQA,A10,90%,0.049,3.867,0.109,5.045,1173.67,33.561,
-HotpotQA,A10,95%,0.07,10.928,0.412,8.659,3079.57,68.626,
-HotpotQA,MAC,85%,0,0,0.974,2.844,2304.125187,34.783,
-HotpotQA,MAC,90%,0,0,1.913,3.542,3415.736201,53.004,
-HotpotQA,MAC,95%,0,0,5.783,6.764,9797.244043,95.413,

research/paper_plot/data/ram_storage.csv

@@ -1,3 +0,0 @@
-Hardware,HNSW,IVF,DiskANN,IVF-Disk,IVF-Recompute,Our,BM25
-RAM,190,171,10,0,0,0,0
-Storage,185.4,171,240,171,0.5,5,59

research/paper_plot/data/swithc_e2e.csv

@@ -1,12 +0,0 @@
-Torch,8,55.592
-Torch,16,75.439
-Torch,32,110.025
-Torch,64,186.496
-Tutel,8,56.718
-Tutel,16,82.121
-Tutel,32,125.070
-Tutel,64,216.191
-BRT,8,56.725
-BRT,16,79.291
-BRT,32,93.180
-BRT,64,118.923

research/paper_plot/data/vary_cache.csv

@@ -1,6 +0,0 @@
-Disk cache size,0,2.5%(180G*2.5%),5%,8%,10%
-Latency,,,,, 
-NQ,4.616,4.133,3.826,3.511,3.323
-TriviaQA,5.777,4.979,4.553,4.141,3.916
-GPQA,1.733,1.593,1.468,1.336,1.259
-Hotpot,15.515,13.479,12.383,11.216,10.606 

research/paper_plot/disk_cache.py

@@ -1,151 +0,0 @@
-import matplotlib
-from matplotlib.axes import Axes
-import numpy as np
-import matplotlib.pyplot as plt
-import matplotlib.patches as mpatches
-from matplotlib.lines import Line2D
-
-# plt.rcParams["font.family"] = "Helvetica"
-plt.rcParams["ytick.direction"] = "in"
-plt.rcParams["hatch.linewidth"] = 1
-plt.rcParams["font.weight"] = "bold"
-plt.rcParams["axes.labelweight"] = "bold"
-plt.rcParams["text.usetex"] = True
-plt.rcParams["font.family"] = "sans-serif"  # Use generic sans-serif family
-plt.rcParams['text.latex.preamble'] = r"""
-\usepackage{helvet} % Use Helvetica font for text
-\usepackage{sfmath} % Use sans-serif font for math
-\renewcommand{\familydefault}{\sfdefault} % Set sans-serif as default text font
-\usepackage[T1]{fontenc} % Recommended for font encoding
-"""
-# plt.rcParams['mathtext.fontset'] = 'dejavusans'
-SAVE_PTH = "./paper_plot/figures"
-font_size = 16
-
-# New data in dictionary format
-datasets = ["NQ", "TriviaQA", "GPQA", "Hotpot"]
-
-cache_ratios = ["4.2G\n (0\%)", "8.7G\n (2.5\%)", "13.2G\n (5\%)", "18.6G\n (8\%)", "22.2G\n (10\%)"]
-latency_data = {
-    "NQ": [4.616, 4.133, 3.826, 3.511, 3.323],
-    "TriviaQA": [5.777, 4.979, 4.553, 4.141, 3.916],
-    "GPQA": [1.733, 1.593, 1.468, 1.336, 1.259],
-    "Hotpot": [15.515, 13.479, 12.383, 11.216, 10.606],
-}
-cache_hit_counts = {
-    "NQ": [0, 14.81, 23.36, 31.99, 36.73],
-    "TriviaQA": [0, 18.55, 27.99, 37.06, 41.86],
-    "GPQA": [0, 10.99, 20.31, 29.71, 35.01],
-    "Hotpot": [0, 17.47, 26.91, 36.2, 41.06]
-}
-
-# Create the figure with 4 subplots in a 2x2 grid
-fig, axes_grid = plt.subplots(2, 2, figsize=(7,6))
-axes = axes_grid.flatten()  # Flatten the 2x2 grid to a 1D array
-
-# Bar style settings
-width = 0.7
-x = np.arange(len(cache_ratios))
-
-# Define hatch patterns for different cache ratios
-hatch_patterns = ['//', '//', '//', '//', '//']
-
-# Find max cache hit value across all datasets for unified y-axis
-all_hit_counts = []
-for dataset in datasets:
-    all_hit_counts.extend(cache_hit_counts[dataset])
-max_unified_hit = max(all_hit_counts) * 1.13
-
-for i, dataset in enumerate(datasets):
-    latencies = latency_data[dataset]
-    hit_counts = cache_hit_counts[dataset]
-
-    for j, val in enumerate(latencies):
-        container = axes[i].bar(
-            x[j],
-            val,
-            width=width,
-            color="white",
-            edgecolor="black",
-            linewidth=1.0,
-            zorder=10,
-        )
-        axes[i].bar_label(
-            container,
-            [f"{val:.2f}"],
-            fontsize=10,
-            zorder=200,
-            fontweight="bold",
-        )
-
-    axes[i].set_title(dataset, fontsize=font_size)
-    axes[i].set_xticks(x)
-    axes[i].set_xticklabels(cache_ratios, fontsize=12, rotation=0, ha='center', fontweight="bold")
-
-    max_val_ratios = [1.35, 1.65, 1.45, 1.75]
-    max_val = max(latencies) * max_val_ratios[i]
-    axes[i].set_ylim(0, max_val)
-    axes[i].tick_params(axis='y', labelsize=12)
-
-    if i % 2 == 0:
-        axes[i].set_ylabel("Latency (s)", fontsize=font_size)
-        axes[i].yaxis.set_major_formatter(matplotlib.ticker.FormatStrFormatter('%.1f'))
-
-    ax2: Axes = axes[i].twinx()
-    ax2.plot(x, hit_counts,
-        linestyle='--',
-        marker='o',
-        markersize=6,
-        linewidth=1.5,
-        color='k',
-        markerfacecolor='none',
-        zorder=20)
-
-    ax2.set_ylim(0, max_unified_hit)
-    ax2.tick_params(axis='y', labelsize=12)
-    if i % 2 == 1:
-        ax2.set_ylabel(r"Cache Hit (\%)", fontsize=font_size)
-
-    for j, val in enumerate(hit_counts):
-        if val > 0:
-            ax2.annotate(f"{val:.1f}%",
-                         (x[j], val),
-                         textcoords="offset points",
-                         xytext=(0, 5),
-                         ha='center',
-                         va='bottom',
-                         fontsize=10,
-                         fontweight='bold')
-
-# Create legend for both plots
-bar_patch = mpatches.Patch(facecolor='white', edgecolor='black', label='Latency')
-line_patch = Line2D([0], [0], color='black', linestyle='--', label='Cache Hit Rate')
-
-# --- MODIFICATION FOR LEGEND AT THE TOP ---
-fig.legend(handles=[bar_patch, line_patch],
-           loc='upper center',        # Position the legend at the upper center
-           bbox_to_anchor=(0.5, 0.995), # Anchor point (0.5 means horizontal center of figure,
-                                      # 0.97 means 97% from the bottom, so near the top)
-           ncol=3,
-           fontsize=font_size-2)
-# --- END OF MODIFICATION ---
-
-# Set common x-axis label - you might want to add this back if needed
-# fig.text(0.5, 0.02, "Disk Cache Size", ha='center', fontsize=font_size, fontweight='bold') # Adjusted y for potential bottom label
-
-# --- MODIFICATION FOR TIGHT LAYOUT ---
-# Adjust rect to make space for the legend at the top.
-# (left, bottom, right, top_for_subplots)
-# We want subplots to occupy space from y=0 up to y=0.93 (or similar)
-# leaving the top portion (0.93 to 1.0) for the legend.
-plt.tight_layout(rect=(0, 0, 1, 0.93)) # Ensure subplots are below the legend
-# --- END OF MODIFICATION ---
-
-# Create directory if it doesn't exist (optional, good practice)
-import os
-if not os.path.exists(SAVE_PTH):
-    os.makedirs(SAVE_PTH)
-
-plt.savefig(f"{SAVE_PTH}/disk_cache_latency.pdf", dpi=300) # Changed filename slightly for testing
-print(f"Save to {SAVE_PTH}/disk_cache_latency.pdf")
-# plt.show() # Optional: to display the plot