docs: Specify core contributors and welcome new contributors

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

@@ -5,7 +5,6 @@ on:
     branches: [ main ]
   pull_request:
     branches: [ main ]
-  workflow_dispatch:
 
 jobs:
   build:

.github/workflows/build-reusable.yml

@@ -54,36 +54,16 @@ jobs:
             python: '3.12'
           - os: ubuntu-22.04
             python: '3.13'
-          - os: macos-14
+          - os: macos-latest
             python: '3.9'
-          - os: macos-14
+          - os: macos-latest
             python: '3.10'
-          - os: macos-14
+          - os: macos-latest
             python: '3.11'
-          - os: macos-14
+          - os: macos-latest
             python: '3.12'
-          - os: macos-14
+          - os: macos-latest
             python: '3.13'
-          - os: macos-15
-            python: '3.9'
-          - os: macos-15
-            python: '3.10'
-          - os: macos-15
-            python: '3.11'
-          - os: macos-15
-            python: '3.12'
-          - os: macos-15
-            python: '3.13'
-          - os: macos-13
-            python: '3.9'
-          - os: macos-13
-            python: '3.10'
-          - os: macos-13
-            python: '3.11'
-          - os: macos-13
-            python: '3.12'
-          # Note: macos-13 + Python 3.13 excluded due to PyTorch compatibility
-          # (PyTorch 2.5+ supports Python 3.13 but not Intel Mac x86_64)
     runs-on: ${{ matrix.os }}
 
     steps:
@@ -129,73 +109,48 @@ jobs:
             uv pip install --system delocate
           fi
 
-      - name: Set macOS environment variables
-        if: runner.os == 'macOS'
-        run: |
-          # Use brew --prefix to automatically detect Homebrew installation path
-          HOMEBREW_PREFIX=$(brew --prefix)
-          echo "HOMEBREW_PREFIX=${HOMEBREW_PREFIX}" >> $GITHUB_ENV
-          echo "OpenMP_ROOT=${HOMEBREW_PREFIX}/opt/libomp" >> $GITHUB_ENV
-
-          # Set CMAKE_PREFIX_PATH to let CMake find all packages automatically
-          echo "CMAKE_PREFIX_PATH=${HOMEBREW_PREFIX}" >> $GITHUB_ENV
-
-          # Set compiler flags for OpenMP (required for both backends)
-          echo "LDFLAGS=-L${HOMEBREW_PREFIX}/opt/libomp/lib" >> $GITHUB_ENV
-          echo "CPPFLAGS=-I${HOMEBREW_PREFIX}/opt/libomp/include" >> $GITHUB_ENV
-
       - name: Build packages
         run: |
           # Build core (platform independent)
-          cd packages/leann-core
+          if [[ "${{ matrix.os }}" == ubuntu-* ]]; then
-          uv build
+            cd packages/leann-core
-          cd ../..
+            uv build
+            cd ../..
+          fi
 
           # Build HNSW backend
           cd packages/leann-backend-hnsw
-          if [[ "${{ matrix.os }}" == macos-* ]]; then
+          if [ "${{ matrix.os }}" == "macos-latest" ]; then
-            # Use system clang for better compatibility
+            # Use system clang instead of homebrew LLVM for better compatibility
             export CC=clang
             export CXX=clang++
-            # Homebrew libraries on each macOS version require matching minimum version
+            export MACOSX_DEPLOYMENT_TARGET=11.0
-            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
+            uv build --wheel --python python
-              export MACOSX_DEPLOYMENT_TARGET=13.0
-            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=14.0
-            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=15.0
-            fi
-            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
           else
-            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
+            uv build --wheel --python python
           fi
           cd ../..
 
           # Build DiskANN backend
           cd packages/leann-backend-diskann
-          if [[ "${{ matrix.os }}" == macos-* ]]; then
+          if [ "${{ matrix.os }}" == "macos-latest" ]; then
-            # Use system clang for better compatibility
+            # 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
-            # But Homebrew libraries on each macOS version require matching minimum version
+            export MACOSX_DEPLOYMENT_TARGET=13.3
-            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
+            uv build --wheel --python python
-              export MACOSX_DEPLOYMENT_TARGET=13.3
-            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=14.0
-            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=15.0
-            fi
-            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
           else
-            uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
+            uv build --wheel --python python
           fi
           cd ../..
 
           # Build meta package (platform independent)
-          cd packages/leann
+          if [[ "${{ matrix.os }}" == ubuntu-* ]]; then
-          uv build
+            cd packages/leann
-          cd ../..
+            uv build
+            cd ../..
+          fi
 
       - name: Repair wheels (Linux)
         if: runner.os == 'Linux'
@@ -221,24 +176,10 @@ jobs:
       - name: Repair wheels (macOS)
         if: runner.os == 'macOS'
         run: |
-          # Determine deployment target based on runner OS
-          # Must match the Homebrew libraries for each macOS version
-          if [[ "${{ matrix.os }}" == "macos-13" ]]; then
-            HNSW_TARGET="13.0"
-            DISKANN_TARGET="13.3"
-          elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
-            HNSW_TARGET="14.0"
-            DISKANN_TARGET="14.0"
-          elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
-            HNSW_TARGET="15.0"
-            DISKANN_TARGET="15.0"
-          fi
-
           # Repair HNSW wheel
           cd packages/leann-backend-hnsw
           if [ -d dist ]; then
-            export MACOSX_DEPLOYMENT_TARGET=$HNSW_TARGET
+            delocate-wheel -w dist_repaired -v dist/*.whl
-            delocate-wheel -w dist_repaired -v --require-target-macos-version $HNSW_TARGET dist/*.whl
             rm -rf dist
             mv dist_repaired dist
           fi
@@ -247,8 +188,7 @@ jobs:
           # Repair DiskANN wheel
           cd packages/leann-backend-diskann
           if [ -d dist ]; then
-            export MACOSX_DEPLOYMENT_TARGET=$DISKANN_TARGET
+            delocate-wheel -w dist_repaired -v dist/*.whl
-            delocate-wheel -w dist_repaired -v --require-target-macos-version $DISKANN_TARGET dist/*.whl
             rm -rf dist
             mv dist_repaired dist
           fi
@@ -259,34 +199,39 @@ jobs:
           echo "📦 Built packages:"
           find packages/*/dist -name "*.whl" -o -name "*.tar.gz" | sort
 
-
       - name: Install built packages for testing
         run: |
-          # Create a virtual environment with the correct Python version
+          # Create a virtual environment
-          uv venv --python ${{ matrix.python }}
+          uv venv
           source .venv/bin/activate || source .venv/Scripts/activate
 
-          # Install packages using --find-links to prioritize local builds
+          # Install the built wheels
-          uv pip install --find-links packages/leann-core/dist --find-links packages/leann-backend-hnsw/dist --find-links packages/leann-backend-diskann/dist packages/leann-core/dist/*.whl || uv pip install --find-links packages/leann-core/dist packages/leann-core/dist/*.tar.gz
+          # Use --find-links to let uv choose the correct wheel for the platform
-          uv pip install --find-links packages/leann-core/dist packages/leann-backend-hnsw/dist/*.whl
+          if [[ "${{ matrix.os }}" == ubuntu-* ]]; then
-          uv pip install --find-links packages/leann-core/dist packages/leann-backend-diskann/dist/*.whl
+            uv pip install leann-core --find-links packages/leann-core/dist
-          uv pip install packages/leann/dist/*.whl || uv pip install packages/leann/dist/*.tar.gz
+            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
+          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
+          PYTORCH_ENABLE_MPS_FALLBACK: 0  # Disable MPS on macOS CI to avoid memory issues
-          OMP_NUM_THREADS: 1
+          OMP_NUM_THREADS: 1  # Disable OpenMP parallelism to avoid libomp crashes
-          MKL_NUM_THREADS: 1
+          MKL_NUM_THREADS: 1  # Single thread for MKL operations
         run: |
+          # Activate virtual environment
           source .venv/bin/activate || source .venv/Scripts/activate
-          pytest tests/ -v --tb=short
+
+          # Run all tests
+          pytest tests/
 
       - name: Run sanity checks (optional)
         run: |

.pre-commit-config.yaml

@@ -1,6 +1,6 @@
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v4.5.0
     hooks:
       - id: trailing-whitespace
       - id: end-of-file-fixer
@@ -10,7 +10,7 @@ repos:
       - id: debug-statements
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.12.7  # Fixed version to match pyproject.toml
+    rev: v0.2.1
     hooks:
       - id: ruff
       - id: ruff-format

README.md

@@ -3,11 +3,9 @@
 </p>
 
 <p align="center">
-  <img src="https://img.shields.io/badge/Python-3.9%20%7C%203.10%20%7C%203.11%20%7C%203.12%20%7C%203.13-blue.svg" alt="Python Versions">
+  <img src="https://img.shields.io/badge/Python-3.9%2B-blue.svg" alt="Python 3.9+">
-  <img src="https://github.com/yichuan-w/LEANN/actions/workflows/build-and-publish.yml/badge.svg" alt="CI Status">
-  <img src="https://img.shields.io/badge/Platform-Ubuntu%20%7C%20macOS%20(ARM64%2FIntel)-lightgrey" alt="Platform">
   <img src="https://img.shields.io/badge/License-MIT-green.svg" alt="MIT License">
-  <img src="https://img.shields.io/badge/MCP-Native%20Integration-blue" alt="MCP Integration">
+  <img src="https://img.shields.io/badge/Platform-Linux%20%7C%20macOS-lightgrey" alt="Platform">
 </p>
 
 <h2 align="center" tabindex="-1" class="heading-element" dir="auto">
@@ -18,10 +16,7 @@ LEANN is an innovative vector database that democratizes personal AI. Transform
 
 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 semantic 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)**, **[codebase](#-claude-code-integration-transform-your-development-workflow)**\* , 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.
-
-
-\* Claude Code only supports basic `grep`-style keyword search. **LEANN** is a drop-in **semantic search MCP service fully compatible with Claude Code**, unlocking intelligent retrieval without changing your workflow. 🔥 Check out [the easy setup →](packages/leann-mcp/README.md)
 
 
 
@@ -31,7 +26,7 @@ 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 text 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)
+> **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)
 
 
 🔒 **Privacy:** Your data never leaves your laptop. No OpenAI, no cloud, no "terms of service".
@@ -71,8 +66,6 @@ source .venv/bin/activate
 uv pip install leann
 ```
 
-> Low-resource? See “Low-resource setups” in the [Configuration Guide](docs/configuration-guide.md#low-resource-setups).
-
 <details>
 <summary>
 <strong>🔧 Build from Source (Recommended for development)</strong>
@@ -173,7 +166,7 @@ ollama pull llama3.2:1b
 
 </details>
 
-### ⭐ Flexible Configuration
+### Flexible Configuration
 
 LEANN provides flexible parameters for embedding models, search strategies, and data processing to fit your specific needs.
 
@@ -186,34 +179,33 @@ All RAG examples share these common parameters. **Interactive mode** is availabl
 
 ```bash
 # Core Parameters (General preprocessing for all examples)
---index-dir DIR              # Directory to store the index (default: current directory)
+--index-dir DIR          # Directory to store the index (default: current directory)
---query "YOUR QUESTION"      # Single query mode. Omit for interactive chat (type 'quit' to exit), and now you can play with your index interactively
+--query "YOUR QUESTION"  # Single query mode. Omit for interactive chat (type 'quit' to exit), and now you can play with your index interactively
---max-items N                # Limit data preprocessing (default: -1, process all data)
+--max-items N           # Limit data preprocessing (default: -1, process all data)
---force-rebuild              # Force rebuild index even if it exists
+--force-rebuild         # Force rebuild index even if it exists
 
 # Embedding Parameters
---embedding-model MODEL      # e.g., facebook/contriever, text-embedding-3-small, mlx-community/Qwen3-Embedding-0.6B-8bit or nomic-embed-text
+--embedding-model MODEL  # e.g., facebook/contriever, text-embedding-3-small or mlx-community/multilingual-e5-base-mlx
---embedding-mode MODE        # sentence-transformers, openai, mlx, or ollama
+--embedding-mode MODE    # sentence-transformers, openai, or mlx
 
 # LLM Parameters (Text generation models)
---llm TYPE                   # LLM backend: openai, ollama, or hf (default: openai)
+--llm TYPE              # LLM backend: openai, ollama, or hf (default: openai)
---llm-model MODEL            # Model name (default: gpt-4o) e.g., gpt-4o-mini, llama3.2:1b, Qwen/Qwen2.5-1.5B-Instruct
+--llm-model MODEL       # Model name (default: gpt-4o) e.g., gpt-4o-mini, llama3.2:1b, Qwen/Qwen2.5-1.5B-Instruct
---thinking-budget LEVEL      # Thinking budget for reasoning models: low/medium/high (supported by o3, o3-mini, GPT-Oss:20b, and other reasoning models)
 
 # Search Parameters
---top-k N                    # Number of results to retrieve (default: 20)
+--top-k N               # Number of results to retrieve (default: 20)
---search-complexity N        # Search complexity for graph traversal (default: 32)
+--search-complexity N   # Search complexity for graph traversal (default: 32)
 
 # Chunking Parameters
---chunk-size N               # Size of text chunks (default varies by source: 256 for most, 192 for WeChat)
+--chunk-size N          # Size of text chunks (default varies by source: 256 for most, 192 for WeChat)
---chunk-overlap N            # Overlap between chunks (default varies: 25-128 depending on source)
+--chunk-overlap N       # Overlap between chunks (default varies: 25-128 depending on source)
 
 # Index Building Parameters
---backend-name NAME          # Backend to use: hnsw or diskann (default: hnsw)
+--backend-name NAME     # Backend to use: hnsw or diskann (default: hnsw)
---graph-degree N             # Graph degree for index construction (default: 32)
+--graph-degree N        # Graph degree for index construction (default: 32)
---build-complexity N         # Build complexity for index construction (default: 64)
+--build-complexity N    # Build complexity for index construction (default: 64)
---compact / --no-compact     # Use compact storage (default: true). Must be `no-compact` for `no-recompute` build.
+--no-compact           # Disable compact index storage (compact storage IS enabled to save storage by default)
---recompute / --no-recompute # Enable/disable embedding recomputation (default: enabled). Should not do a `no-recompute` search in a `recompute` build.
+--no-recompute         # Disable embedding recomputation (recomputation IS enabled to save storage by default)
 ```
 
 </details>
@@ -226,7 +218,7 @@ Ask questions directly about your personal PDFs, documents, and any directory co
   <img src="videos/paper_clear.gif" alt="LEANN Document Search Demo" width="600">
 </p>
 
-The example below asks a question about summarizing our paper (uses default data in `data/`, which is a directory with diverse data sources: two papers, Pride and Prejudice, and a Technical report about LLM in Huawei in Chinese), and this is the **easiest example** to run here:
+The example below asks a question about summarizing our paper (uses default data in `data/`, which is a directory with diverse data sources: two papers, Pride and Prejudice, and a README in Chinese) and this is the **easiest example** to run here:
 
 ```bash
 source .venv/bin/activate # Don't forget to activate the virtual environment
@@ -421,26 +413,7 @@ Once the index is built, you can ask questions like:
 
 </details>
 
-### 🚀 Claude Code Integration: Transform Your Development Workflow!
 
-**The future of code assistance is here.** Transform your development workflow with LEANN's native MCP integration for Claude Code. Index your entire codebase and get intelligent code assistance directly in your IDE.
-
-**Key features:**
-- 🔍 **Semantic code search** across your entire project
-- 📚 **Context-aware assistance** for debugging and development
-- 🚀 **Zero-config setup** with automatic language detection
-
-```bash
-# Install LEANN globally for MCP integration
-uv tool install leann-core
-
-# Setup is automatic - just start using Claude Code!
-```
-Try our fully agentic pipeline with auto query rewriting, semantic search planning, and more:
-
-![LEANN MCP Integration](assets/mcp_leann.png)
-
-**Ready to supercharge your coding?** [Complete Setup Guide →](packages/leann-mcp/README.md)
 
 ## 🖥️ Command Line Interface
 
@@ -454,24 +427,22 @@ source .venv/bin/activate
 leann --help
 ```
 
-**To make it globally available:**
+**To make it globally available (recommended for daily use):**
 ```bash
 # Install the LEANN CLI globally using uv tool
-uv tool install leann-core
+uv tool install leann
 
 # Now you can use leann from anywhere without activating venv
 leann --help
 ```
 
-> **Note**: Global installation is required for Claude Code integration. The `leann_mcp` server depends on the globally available `leann` command.
-
 
 
 ### Usage Examples
 
 ```bash
-# build from a specific directory, and my_docs is the index name(Here you can also build from multiple dict or multiple files)
+# Build an index from documents
-leann build my-docs --docs ./your_documents
+leann build my-docs --docs ./documents
 
 # Search your documents
 leann search my-docs "machine learning concepts"
@@ -484,29 +455,27 @@ leann list
 ```
 
 **Key CLI features:**
-- Auto-detects document formats (PDF, TXT, MD, DOCX, PPTX + code files)
+- 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/` (project-local)
+- Organized index storage in `~/.leann/indexes/`
 - Support for advanced search parameters
 
 <details>
 <summary><strong>📋 Click to expand: Complete CLI Reference</strong></summary>
 
-You can use `leann --help`, or `leann build --help`, `leann search --help`, `leann ask --help` to get the complete CLI reference.
-
 **Build Command:**
 ```bash
-leann build INDEX_NAME --docs DIRECTORY|FILE [DIRECTORY|FILE ...] [OPTIONS]
+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)
+  --graph-degree N            Graph degree (default: 32)
-  --complexity N               Build complexity (default: 64)
+  --complexity N              Build complexity (default: 64)
-  --force                      Force rebuild existing index
+  --force                     Force rebuild existing index
-  --compact / --no-compact     Use compact storage (default: true). Must be `no-compact` for `no-recompute` build.
+  --compact                   Use compact storage (default: true)
-  --recompute / --no-recompute Enable recomputation (default: true)
+  --recompute                 Enable recomputation (default: true)
 ```
 
 **Search Command:**
@@ -514,9 +483,9 @@ Options:
 leann search INDEX_NAME QUERY [OPTIONS]
 
 Options:
-  --top-k N                     Number of results (default: 5)
+  --top-k N                   Number of results (default: 5)
-  --complexity N                Search complexity (default: 64)
+  --complexity N              Search complexity (default: 64)
-  --recompute / --no-recompute  Enable/disable embedding recomputation (default: enabled). Should not do a `no-recompute` search in a `recompute` build.
+  --recompute-embeddings      Use recomputation for highest accuracy
   --pruning-strategy {global,local,proportional}
 ```
 
@@ -547,16 +516,12 @@ Options:
 - **Dynamic batching:** Efficiently batch embedding computations for GPU utilization
 - **Two-level search:** Smart graph traversal that prioritizes promising nodes
 
-**Backends:**
+**Backends:** HNSW (default) for most use cases, with optional DiskANN support for billion-scale datasets.
-- **HNSW** (default): Ideal for most datasets with maximum storage savings through full recomputation
-- **DiskANN**: Advanced option with superior search performance, using PQ-based graph traversal with real-time reranking for the best speed-accuracy trade-off
 
 ## Benchmarks
 
-**[DiskANN vs HNSW Performance Comparison →](benchmarks/diskann_vs_hnsw_speed_comparison.py)** - Compare search performance between both backends
-
-**[Simple Example: Compare LEANN vs FAISS →](benchmarks/compare_faiss_vs_leann.py)** - See storage savings in action
 
+**[Simple Example: Compare LEANN vs FAISS →](benchmarks/compare_faiss_vs_leann.py)**
 ### 📊 Storage Comparison
 
 | System | DPR (2.1M) | Wiki (60M) | Chat (400K) | Email (780K) | Browser (38K) |
@@ -615,9 +580,8 @@ We welcome more contributors! Feel free to open issues or submit PRs.
 
 This work is done at [**Berkeley Sky Computing Lab**](https://sky.cs.berkeley.edu/).
 
-## Star History
+---
 
-[![Star History Chart](https://api.star-history.com/svg?repos=yichuan-w/LEANN&type=Date)](https://www.star-history.com/#yichuan-w/LEANN&Date)
 <p align="center">
   <strong>⭐ Star us on GitHub if Leann is useful for your research or applications!</strong>
 </p>

apps/base_rag_example.py

@@ -69,14 +69,14 @@ class BaseRAGExample(ABC):
             "--embedding-model",
             type=str,
             default=embedding_model_default,
-            help=f"Embedding model to use (default: {embedding_model_default}), we provide facebook/contriever, text-embedding-3-small,mlx-community/Qwen3-Embedding-0.6B-8bit or nomic-embed-text",
+            help=f"Embedding model to use (default: {embedding_model_default})",
         )
         embedding_group.add_argument(
             "--embedding-mode",
             type=str,
             default="sentence-transformers",
-            choices=["sentence-transformers", "openai", "mlx", "ollama"],
+            choices=["sentence-transformers", "openai", "mlx"],
-            help="Embedding backend mode (default: sentence-transformers), we provide sentence-transformers, openai, mlx, or ollama",
+            help="Embedding backend mode (default: sentence-transformers)",
         )
 
         # LLM parameters
@@ -85,14 +85,14 @@ class BaseRAGExample(ABC):
             "--llm",
             type=str,
             default="openai",
-            choices=["openai", "ollama", "hf", "simulated"],
+            choices=["openai", "ollama", "hf"],
-            help="LLM backend: openai, ollama, or hf (default: openai)",
+            help="LLM backend to use (default: openai)",
         )
         llm_group.add_argument(
             "--llm-model",
             type=str,
             default=None,
-            help="Model name (default: gpt-4o) e.g., gpt-4o-mini, llama3.2:1b, Qwen/Qwen2.5-1.5B-Instruct",
+            help="LLM model name (default: gpt-4o for openai, llama3.2:1b for ollama)",
         )
         llm_group.add_argument(
             "--llm-host",
@@ -100,13 +100,6 @@ class BaseRAGExample(ABC):
             default="http://localhost:11434",
             help="Host for Ollama API (default: http://localhost:11434)",
         )
-        llm_group.add_argument(
-            "--thinking-budget",
-            type=str,
-            choices=["low", "medium", "high"],
-            default=None,
-            help="Thinking budget for reasoning models (low/medium/high). Supported by GPT-Oss:20b and other reasoning models.",
-        )
 
         # Search parameters
         search_group = parser.add_argument_group("Search Parameters")
@@ -178,9 +171,6 @@ class BaseRAGExample(ABC):
             config["host"] = args.llm_host
         elif args.llm == "hf":
             config["model"] = args.llm_model or "Qwen/Qwen2.5-1.5B-Instruct"
-        elif args.llm == "simulated":
-            # Simulated LLM doesn't need additional configuration
-            pass
 
         return config
 
@@ -238,17 +228,7 @@ class BaseRAGExample(ABC):
                 if not query:
                     continue
 
-                # Prepare LLM kwargs with thinking budget if specified
+                response = chat.ask(query, top_k=args.top_k, complexity=args.search_complexity)
-                llm_kwargs = {}
-                if hasattr(args, "thinking_budget") and args.thinking_budget:
-                    llm_kwargs["thinking_budget"] = args.thinking_budget
-
-                response = chat.ask(
-                    query,
-                    top_k=args.top_k,
-                    complexity=args.search_complexity,
-                    llm_kwargs=llm_kwargs,
-                )
                 print(f"\nAssistant: {response}\n")
 
             except KeyboardInterrupt:
@@ -267,15 +247,7 @@ class BaseRAGExample(ABC):
         )
 
         print(f"\n[Query]: \033[36m{query}\033[0m")
-
+        response = chat.ask(query, top_k=args.top_k, complexity=args.search_complexity)
-        # Prepare LLM kwargs with thinking budget if specified
-        llm_kwargs = {}
-        if hasattr(args, "thinking_budget") and args.thinking_budget:
-            llm_kwargs["thinking_budget"] = args.thinking_budget
-
-        response = chat.ask(
-            query, top_k=args.top_k, complexity=args.search_complexity, llm_kwargs=llm_kwargs
-        )
         print(f"\n[Response]: \033[36m{response}\033[0m")
 
     async def run(self):

benchmarks/README.md

@@ -1,24 +1,9 @@
-# 🧪 LEANN Benchmarks & Testing
+# 🧪 Leann Sanity Checks
 
-This directory contains performance benchmarks and comprehensive tests for the LEANN system, including backend comparisons and sanity checks across different configurations.
+This directory contains comprehensive sanity checks for the Leann system, ensuring all components work correctly across different configurations.
 
 ## 📁 Test Files
 
-### `diskann_vs_hnsw_speed_comparison.py`
-Performance comparison between DiskANN and HNSW backends:
-- ✅ **Search latency** comparison with both backends using recompute
-- ✅ **Index size** and **build time** measurements
-- ✅ **Score validity** testing (ensures no -inf scores)
-- ✅ **Configurable dataset sizes** for different scales
-
-```bash
-# Quick comparison with 500 docs, 10 queries
-python benchmarks/diskann_vs_hnsw_speed_comparison.py
-
-# Large-scale comparison with 2000 docs, 20 queries
-python benchmarks/diskann_vs_hnsw_speed_comparison.py 2000 20
-```
-
 ### `test_distance_functions.py`
 Tests all supported distance functions across DiskANN backend:
 - ✅ **MIPS** (Maximum Inner Product Search)

benchmarks/benchmark_no_recompute.py

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

benchmarks/diskann_vs_hnsw_speed_comparison.py

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

data/huawei_pangu.md

@@ -1,82 +0,0 @@
-# 盘古之殇：华为诺亚盘古大模型研发历程的心酸与黑暗
-
-各位好，
-
-我是一名盘古大模型团队，华为诺亚方舟实验室的员工。
-
-首先为自证身份，列举一些细节：
-
-1. 现诺亚主任，前算法应用部部长，后改名为小模型实验室的主任王云鹤。前诺亚主任：姚骏（大家称姚老师）。几个实验室主任：唐睿明（明哥，明队，已离职），尚利峰，张维（维哥），郝建业（郝老师），刘武龙（称呼为武龙所）等。其他骨干成员和专家陆续有很多人离职。
-2. 我们隶属于“四野”这个组织。四野下属有许多纵队，基础语言大模型是四纵。王云鹤的小模型是十六纵队。我们参加过苏州的集结，有各种月份的时间节点。在苏州攻关会颁发任务令，需要在节点前达成目标。苏州集结会把各地的人员都集中在苏州研究所，平常住宾馆，比如在甪直的酒店，与家人孩子天各一方。
-3. 在苏州集结的时候周六默认上班，非常辛苦，不过周六有下午茶，有一次还有小龙虾。在苏州研究所的工位搬迁过一次，从一栋楼换到了另一栋。苏州研究所楼栋都是欧式装修，门口有大坡，里面景色很不错。去苏州集结一般至少要去一周，甚至更久，多的人甚至一两个月都回不了家。
-4. 诺亚曾经传说是研究型的，但是来了之后因为在四野做大模型项目，项目成员完全变成了交付型的，且充满了例会，评审，汇报。很多时候做实验都要申请。团队需要对接终端小艺，华为云，ICT等诸多业务线，交付压力不小。
-5. 诺亚研发的盘古模型早期内部代号叫做“盘古智子”，一开始只有内部需要申请试用的网页版，到后续迫于压力在welink上接入和公测开放。
-
-这些天发生关于质疑盘古大模型抄袭千问的事情闹的沸沸扬扬。作为一个盘古团队的成员，我最近夜夜辗转反侧，难以入眠。盘古的品牌受到如此大的影响，一方面，我自私的为我的职业发展担忧，也为自己过去的努力工作感到不值。另一方面，由于有人开始揭露这些事情我内心又感到大快人心。在多少个日日夜夜，我们对内部某些人一次次靠着造假而又获得了无数利益的行为咬牙切齿而又无能为力。这种压抑和羞辱也逐渐消磨了我对华为的感情，让我在这里的时日逐渐浑浑噩噩，迷茫无措，时常怀疑自己的人生和自我价值。
-
-我承认我是一个懦弱的人，作为一个小小的打工人，我不仅不敢和王云鹤等内部手眼通天的人做对，更不敢和华为这样的庞然大物做对。我很怕失去我的工作，毕竟我也有家人和孩子，所以我打心眼里很佩服揭露者。但是，看到内部还在试图洗地掩盖事实，蒙蔽公众的时候，我实在不能容忍了。我也希望勇敢一次，顺从自己本心。就算自损八百，我也希望能伤敌一千。我决定把我在这里的所见所闻（部分来自于同事口述）公布出来，关于盘古大模型的“传奇故事”：
-
-华为确实主要在昇腾卡上训练大模型（小模型实验室有不少英伟达的卡，他们之前也会用来训练，后面转移到昇腾）。曾经我被华为“打造世界第二选择”的决心而折服，我本身也曾经对华为有深厚的感情。我们陪着昇腾一步步摸爬滚打，从充满bug到现在能训出模型，付出了巨大的心血和代价。
-
-最初我们的算力非常有限，在910A上训练模型。那会只支持fp16，训练的稳定性远不如bf16。盘古的moe开始很早，23年就主要是训练38Bmoe模型和后续的71B dense模型。71B的dense模型通过扩增变成了第一代的135Bdense模型，后面主力模型也逐渐在910B上训练。
-
-71B和135B模型都有一个巨大的硬伤就是tokenizer。当时使用的tokenizer编码效率极低，每个单个的符号，数字，空格，乃至汉字都会占用一个token。可想而知这会非常浪费算力，且使得模型的效果很差。这时候小模型实验室正好有个自己训的词表。姚老师当时怀疑是不是模型的tokenizer不好（虽然事后来看，他的怀疑是无疑正确的），于是就决定，让71B和135B换tokenizer，因为小模型实验室曾经尝试过。团队缝合了两个tokenizer，开始了tokenizer的更换。71B模型的更换失败了，而135B因为采用了更精细的embedding初始化策略，续训了至少1T的数据后词表总算更换成功，但可想而知，效果并不会变好。
-
-于此同期，阿里和智谱等国内其他公司在GPU上训练，且已经摸索出了正确的方法，盘古和竞品的差距越来越大。内部一个230B从头训练的dense模型又因为各种原因训练失败，导致项目的状况几乎陷入绝境。面临几个节点的压力以及内部对盘古的强烈质疑时，团队的士气低迷到了极点。团队在算力极其有限的时候，做出了很多努力和挣扎。比如，团队偶然发现当时的38B moe并没有预期moe的效果。于是去掉了moe参数，还原为了13B的dense模型。由于38B的moe源自很早的pangu alpha 13B，架构相对落后，团队进行了一系列的操作，比如切换绝对位置编码到rope，去掉bias，切换为rmsnorm。同时鉴于tokenizer的一些失败和换词表的经验，这个模型的词表也更换为了王云鹤的小模型实验室7B模型所使用的词表。后面这个13B模型进行了扩增续训，变成了第二代38B dense模型（在几个月内这个模型都是主要的盘古中档位模型），曾经具有一定的竞争力。但是，由于更大的135B模型架构落后，且更换词表模型损伤巨大（后续分析发现当时更换的缝合词表有更严重的bug），续训后也与千问等当时国内领先模型存在很大差距。这时由于内部的质疑声和领导的压力也越来越大。团队的状态几乎陷入了绝境。
-
-在这种情况下，王云鹤和他的小模型实验室出手了。他们声称是从旧的135B参数继承改造而来，通过训练短短的几百B数据，各项指标平均提升了十个点左右。实际上，这就是他们套壳应用到大模型的第一次杰作。华为的外行领导内行，使得领导完全对于这种扯淡的事情没有概念，他们只会觉得肯定是有什么算法创新。经过内部的分析，他们实际上是使用Qwen 1.5 110B续训而来，通过加层，扩增ffn维度，添加盘古pi论文的一些机制得来，凑够了大概135B的参数。实际上，旧的135B有107层，而这个模型只有82层，各种配置也都不一样。新的来路不明的135B训练完很多参数的分布也和Qwen 110B几乎一模一样。连模型代码的类名当时都是Qwen，甚至懒得改名。后续这个模型就是所谓的135B V2。而这个模型当时也提供给了很多下游，甚至包括外部客户。
-
-这件事对于我们这些认真诚实做事的同事们带来了巨大的冲击，内部很多人其实都知道这件事，甚至包括终端和华为云。我们都戏称以后别叫盘古模型了，叫千古吧。当时团队成员就想向bcg举报了，毕竟这已经是重大的业务造假了。但是后面据说被领导拦了下来，因为更高级别的领导（比如姚老师，以及可能熊总和查老）其实后面也知道了，但是并不管，因为通过套壳拿出好的结果，对他们也是有利的。这件事使得当时团队几位最强的同事开始心灰意冷，离职跑路也逐渐成为挂在嘴边的事。
-
-此时，盘古似乎迎来了转机。由于前面所述的这些盘古模型基本都是续训和改造而来，当时诺亚完全没有掌握从头训练的技术，何况还是在昇腾的NPU上进行训练。在当时团队的核心成员的极力争取下，盘古开始了第三代模型的训练，付出了巨大的努力后，在数据架构和训练算法方面都与业界逐渐接轨，而这其中的艰辛和小模型实验室的人一点关系都没有。
-
-一开始团队成员毫无信心，只从一个13B的模型开始训练，但是后面发现效果还不错，于是这个模型后续再次进行了一次参数扩增，变成了第三代的38B，代号38B V3。想必很多产品线的兄弟都对这个模型很熟悉。当时这个模型的tokenizer是基于llama的词表进行扩展的（也是业界常见的做法）。而当时王云鹤的实验室做出来了另一个词表（也就是后续pangu系列的词表）。当时两个词表还被迫进行了一次赛马，最终没有明显的好坏结论。于是，领导当即决定，应该统一词表，使用王云鹤他们的。于是，在后续从头训练的135B V3（也就是对外的Pangu Ultra），便是采用了这个tokenizer。这也解释了很多使用我们模型的兄弟的疑惑，为什么当时同为V3代的两个不同档位的模型，会使用不同的tokenizer。
-
-
-我们打心眼里觉得，135B V3是我们四纵团队当时的骄傲。这是第一个真正意义上的，华为全栈自研，正经从头训练的千亿级别的模型，且效果与24年同期竞品可比的。写到这里我已经热泪盈眶，太不容易了。当时为了稳定训练，团队做了大量实验对比，并且多次在模型梯度出现异常的时候进行及时回退重启。这个模型真正做到了后面技术报告所说的训练全程没有一个loss spike。我们克服了不知道多少困难，我们做到了，我们愿用生命和荣誉保证这个模型训练的真实性。多少个凌晨，我们为了它的训练而不眠。在被内部心声骂的一文不值的时候，我们有多么不甘，有多少的委屈，我们挺住了。
-
-我们这帮人是真的在为打磨国产算力底座燃烧自己的青春啊……客居他乡，我们放弃了家庭，放弃了假期，放弃了健康，放弃了娱乐，抛头颅洒热血，其中的艰辛与困苦，寥寥数笔不足以概括其万一。在各种动员大会上，当时口号中喊出的盘古必胜，华为必胜，我们心里是真的深深被感动。
-
-然而，我们的所有辛苦的成果，经常被小模型实验室轻飘飘的拿走了。数据，直接要走。代码，直接要走，还要求我们配合适配到能一键运行。我们当时戏称小模型实验室为点鼠标实验室。我们付出辛苦，他们取得荣耀。果然应了那句话，你在负重前行是因为有人替你岁月静好。在这种情况下，越来越多的战友再也坚持不下去了，选择了离开。看到身边那些优秀的同事一个个离职，我的内心又感叹又难过。在这种作战一样的环境下，我们比起同事来说更像是战友。他们在技术上也有无数值得我学习的地方，堪称良师。看到他们去了诸如字节Seed，Deepseek，月之暗面，腾讯和快手等等很多出色的团队，我打心眼里为他们高兴和祝福，脱离了这个辛苦却肮脏的地方。我至今还对一位离职同事的话记忆犹新，ta说：“来这里是我技术生涯中的耻辱，在这里再呆每一天都是浪费生命”。话虽难听却让我无言以对。我担心我自己技术方面的积累不足，以及没法适应互联网公司高淘汰的环境，让我多次想离职的心始终没有迈出这一步。
-
-盘古除了dense模型，后续也启动了moe的探索。一开始训练的是一个224B的moe模型。而与之平行的，小模型实验室也开启了第二次主要的套壳行动（次要的插曲可能还包括一些别的模型，比如math模型），即这次流传甚广的pangu pro moe 72B。这个模型内部自称是从小模型实验室的7B扩增上来的（就算如此，这也与技术报告不符，何况是套壳qwen 2.5的14b续训）。还记得他们训了没几天，内部的评测就立刻追上了当时的38B V3。AI系统实验室很多兄弟因为需要适配模型，都知道他们的套壳行动，只是迫于各种原因，无法伸张正义。实际上，对于后续训了很久很久的这个模型，Honestagi能够分析出这个量级的相似性我已经很诧异了，因为这个模型为了续训洗参数，所付出的算力甚至早就足够从头训一个同档位的模型了。听同事说他们为了洗掉千问的水印，采取了不少办法，甚至包括故意训了脏数据。这也为学术界研究模型血缘提供了一个前所未有的特殊模范吧。以后新的血缘方法提出可以拿出来溜溜。
-
-24年底和25年初，在Deepseek v3和r1发布之后，由于其惊艳的技术水平，团队受到了巨大的冲击，也受到了更大的质疑。于是为了紧跟潮流，盘古模仿Deepseek的模型尺寸，开启了718B moe的训练。这个时候，小模型实验室再次出手了。他们选择了套壳Deepseekv3续训。他们通过冻住Deepseek加载的参数，进行训练。连任务加载ckpt的目录都是deepseekv3，改都不改，何其嚣张？与之相反，一些有真正技术信仰的同事，在从头训练另一个718B的moe。但其中出现了各种各样的问题。但是很显然，这个模型怎么可能比直接套壳的好呢？如果不是团队leader坚持，早就被叫停了。
-
-华为的流程管理之繁重，严重拖累了大模型的研发节奏，例如版本管理，模型血缘，各种流程化，各种可追溯。讽刺的是，小模型实验室的模型似乎从来不受这些流程的约束，想套壳就套壳，想续训就续训，算力源源不断的伸手拿走。这种强烈到近乎魔幻的对比，说明了当前流程管理的情况：只许州官放火，不许百姓点灯。何其可笑？何其可悲？何其可恶？何其可耻！
-
-HonestAGI的事情出来后，内部让大家不停的研讨分析，如何公关和“回应”。诚然，这个原文的分析也许不够有力，给了王云鹤与小模型实验室他们狡辩和颠倒黑白的机会。为此，这两天我内心感到作呕，时时怀疑自己的人生意义以及苍天无眼。我不奉陪了，我要离职了，同时我也在申请从盘古部分技术报告的作者名单中移除。曾经在这些技术报告上署名是我一生都无法抹除的污点。当时我没想到，他们竟然猖狂到敢开源。我没想到，他们敢如此愚弄世人，大肆宣发。当时，我也许是存了侥幸心理，没有拒绝署名。我相信很多扎实做事的战友，也只是被迫上了贼船，或者不知情。但这件事已经无法挽回，我希望我的余生能够坚持扎实做真正有意义的事，为我当时的软弱和不坚定赎罪。
-
-深夜写到这里，我已经泪流满面，泣不成声。还记得一些出色的同事离职时，我苦笑问他们要不要发个长长的心声惯例帖，揭露一下现状。对方说：不了，浪费时间，而且我也怕揭露出来你们过的更糟。我当时一下黯然神伤，因为曾经共同为了理想奋斗过的战友已经彻底对华为彻底灰心了。当时大家调侃，我们用着当年共产党的小米加步枪，组织却有着堪比当年国民党的作风。
-
-曾几何时，我为我们用着小米加步枪打败洋枪洋炮而自豪。
-
-现在，我累了，我想投降。
-
-其实时至今日，我还是真心希望华为能认真吸取教训，能做好盘古，把盘古做到世界一流，把昇腾变成英伟达的水平。内部的劣币驱逐良币，使得诺亚乃至华为在短时间内急剧流失了大量出色的大模型人才。相信他们也正在如Deepseek等各个团队闪耀着，施展着他们的抱负才华，为中美在AI的激烈竞赛中奉献力量。我时常感叹，华为不是没有人才，而是根本不知道怎么留住人才。如果给这些人合适的环境，合适的资源，更少的枷锁，更少的政治斗争，盘古何愁不成？
-
-最后：我以生命，人格和荣誉发誓，我写的以上所有内容均为真实（至少在我有限的认知范围内）。我没有那么高的技术水平以及机会去做详尽扎实的分析，也不敢直接用内部记录举证，怕因为信息安全抓到。但是我相信我很多曾经的战友，会为我作证。在华为内部的兄弟，包括我们曾经服务过的产品线兄弟们，相信本文的无数细节能和你们的印象对照，印证我的说法。你们可能也曾经被蒙骗，但这些残酷的真相不会被尘封。我们奋战过的痕迹，也不应该被扭曲和埋葬。
-
-写了这么多，某些人肯定想把我找出来，抹杀掉。公司搞不好也想让我噤声乃至追责。如果真的这样，我，乃至我的家人的人身乃至生命安全可能都会受到威胁。为了自我保护，我近期每天会跟大家报平安。
-
-如果我消失了，就当是我为了真理和理想，为了华为乃至中国能够更好地发展算力和AI而牺牲了吧，我愿埋葬于那片曾经奋斗过的地方。
-
-诺亚，再见
-
-2025年7月6日凌晨      写于深圳
-
----
-
-各位好，
-
-感谢大家的关心与祝福。我目前暂时安全，但公司应该在进行排查与某些名单收集，后续情况未知。
-
-我补充一些细节，以免某些人继续颠倒黑白。
-
-关于135B V2，小模型实验室在迅速地完成套壳并拿完所有套壳带来的好处后（比如任务令表彰和及时激励），因为不想继续支撑下游应用和模型迭代，又把这个烫手山芋甩给了四纵。确实技高一筹，直接把四纵的兄弟们拉下水。同事提供过去一个老旧的模型，最终拿回了一个当时一个魔改的先进的千问。做大模型的人，自己做的模型就像自己孩子一样熟悉，不要把别人都当傻子。就像自家儿子出门一趟，回来个别人家孩子。
-
-盘古report的署名是不符合学术规范的。例如，135B V3有不少有技术贡献的人，因为作者名额数量限制，劳动成果没有得到应有的回报，团队内曾经有不小的意见。这个模型当时是大家智慧和汗水的结晶，甚至是团队当时的精神支柱，支撑着不少兄弟们继续留在诺亚。所谓的名额限制，以及挂名了一些毫无技术贡献的人（如一些小模型实验室的人），让兄弟们何其心寒。
-
----
-
-暂时平安。另外，支持我勇于说出真相的战友们 https://github.com/HW-whistleblower/True-Story-of-Pangu/issues/317

docs/THINKING_BUDGET_FEATURE.md

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

docs/configuration-guide.md

@@ -49,25 +49,14 @@ Based on our experience developing LEANN, embedding models fall into three categ
 - **Cons**: Slower inference, longer index build times
 - **Use when**: Quality is paramount and you have sufficient compute resources. **Highly recommended** for production use
 
-### Quick Start: Cloud and Local Embedding Options
+### Quick Start: OpenAI Embeddings (Fastest Setup)
 
-**OpenAI Embeddings (Fastest Setup)**
+For immediate testing without local model downloads:
-For immediate testing without local model downloads(also if you [do not have GPU](https://github.com/yichuan-w/LEANN/issues/43) and do not care that much about your document leak, you should use this, we compute the embedding and recompute using openai API):
 ```bash
 # Set OpenAI embeddings (requires OPENAI_API_KEY)
 --embedding-mode openai --embedding-model text-embedding-3-small
 ```
 
-**Ollama Embeddings (Privacy-Focused)**
-For local embeddings with complete privacy:
-```bash
-# First, pull an embedding model
-ollama pull nomic-embed-text
-
-# Use Ollama embeddings
---embedding-mode ollama --embedding-model nomic-embed-text
-```
-
 <details>
 <summary><strong>Cloud vs Local Trade-offs</strong></summary>
 
@@ -97,24 +86,16 @@ ollama pull nomic-embed-text
 ```
 
 ### DiskANN
-**Best for**: Large datasets, especially when you want `recompute=True`.
+**Best for**: Large datasets (> 10M vectors, 10GB+ index size) - **⚠️ Beta version, still in active development**
-
+- Uses Product Quantization (PQ) for coarse filtering during graph traversal
-**Key advantages:**
+- Novel approach: stores only PQ codes, performs rerank with exact computation in final step
-- **Faster search** on large datasets (3x+ speedup vs HNSW in many cases)
+- Implements a corner case of double-queue: prunes all neighbors and recomputes at the end
-- **Smart storage**: `recompute=True` enables automatic graph partitioning for smaller indexes
-- **Better scaling**: Designed for 100k+ documents
-
-**Recompute behavior:**
-- `recompute=True` (recommended): Pure PQ traversal + final reranking - faster and enables partitioning
-- `recompute=False`: PQ + partial real distances during traversal - slower but higher accuracy
 
 ```bash
-# Recommended for most use cases
+# For billion-scale deployments
---backend-name diskann --graph-degree 32 --build-complexity 64
+--backend-name diskann --graph-degree 64 --build-complexity 128
 ```
 
-**Performance Benchmark**: Run `uv run benchmarks/diskann_vs_hnsw_speed_comparison.py` to compare DiskANN and HNSW on your system.
-
 ## LLM Selection: Engine and Model Comparison
 
 ### LLM Engines
@@ -122,15 +103,13 @@ ollama pull nomic-embed-text
 **OpenAI** (`--llm openai`)
 - **Pros**: Best quality, consistent performance, no local resources needed
 - **Cons**: Costs money ($0.15-2.5 per million tokens), requires internet, data privacy concerns
-- **Models**: `gpt-4o-mini` (fast, cheap), `gpt-4o` (best quality), `o3` (reasoning), `o3-mini` (reasoning, cheaper)
+- **Models**: `gpt-4o-mini` (fast, cheap), `gpt-4o` (best quality), `o3-mini` (reasoning, not so expensive)
-- **Thinking Budget**: Use `--thinking-budget low/medium/high` for o-series reasoning models (o3, o3-mini, o4-mini)
 - **Note**: Our current default, but we recommend switching to Ollama for most use cases
 
 **Ollama** (`--llm ollama`)
 - **Pros**: Fully local, free, privacy-preserving, good model variety
 - **Cons**: Requires local GPU/CPU resources, slower than cloud APIs, need to install extra [ollama app](https://github.com/ollama/ollama?tab=readme-ov-file#ollama) and pre-download models by `ollama pull`
 - **Models**: `qwen3:0.6b` (ultra-fast), `qwen3:1.7b` (balanced), `qwen3:4b` (good quality), `qwen3:7b` (high quality), `deepseek-r1:1.5b` (reasoning)
-- **Thinking Budget**: Use `--thinking-budget low/medium/high` for reasoning models like GPT-Oss:20b
 
 **HuggingFace** (`--llm hf`)
 - **Pros**: Free tier available, huge model selection, direct model loading (vs Ollama's server-based approach)
@@ -172,36 +151,6 @@ ollama pull nomic-embed-text
 - LLM processing time ∝ top_k × chunk_size
 - Total context = top_k × chunk_size tokens
 
-### Thinking Budget for Reasoning Models
-
-**`--thinking-budget`** (reasoning effort level)
-- Controls the computational effort for reasoning models
-- Options: `low`, `medium`, `high`
-- Guidelines:
-  - `low`: Fast responses, basic reasoning (default for simple queries)
-  - `medium`: Balanced speed and reasoning depth
-  - `high`: Maximum reasoning effort, best for complex analytical questions
-- **Supported Models**:
-  - **Ollama**: `gpt-oss:20b`, `gpt-oss:120b`
-  - **OpenAI**: `o3`, `o3-mini`, `o4-mini`, `o1` (o-series reasoning models)
-- **Note**: Models without reasoning support will show a warning and proceed without reasoning parameters
-- **Example**: `--thinking-budget high` for complex analytical questions
-
-**📖 For detailed usage examples and implementation details, check out [Thinking Budget Documentation](THINKING_BUDGET_FEATURE.md)**
-
-**💡 Quick Examples:**
-```bash
-# OpenAI o-series reasoning model
-python apps/document_rag.py --query "What are the main techniques LEANN explores?" \
-  --index-dir hnswbuild --backend hnsw \
-  --llm openai --llm-model o3 --thinking-budget medium
-
-# Ollama reasoning model
-python apps/document_rag.py --query "What are the main techniques LEANN explores?" \
-  --index-dir hnswbuild --backend hnsw \
-  --llm ollama --llm-model gpt-oss:20b --thinking-budget high
-```
-
 ### Graph Degree (HNSW/DiskANN)
 
 **`--graph-degree`**
@@ -230,15 +179,9 @@ python apps/document_rag.py --query "What are the main techniques LEANN explores
 
 3. **Use MLX on Apple Silicon** (optional optimization):
    ```bash
-   --embedding-mode mlx --embedding-model mlx-community/Qwen3-Embedding-0.6B-8bit
+   --embedding-mode mlx --embedding-model mlx-community/multilingual-e5-base-mlx
    ```
-    MLX might not be the best choice, as we tested and found that it only offers 1.3x acceleration compared to HF, so maybe using ollama is a better choice for embedding generation
 
-4. **Use Ollama**
-   ```bash
-   --embedding-mode ollama --embedding-model nomic-embed-text
-   ```
-   To discover additional embedding models in ollama, check out https://ollama.com/search?c=embedding or read more about embedding models at https://ollama.com/blog/embedding-models, please do check the model size that works best for you
 ### If Search Quality is Poor
 
 1. **Increase retrieval count**:
@@ -267,118 +210,27 @@ Every configuration choice involves trade-offs:
 
 The key is finding the right balance for your specific use case. Start small and simple, measure performance, then scale up only where needed.
 
-## Low-resource setups
+## Deep Dive: Critical Configuration Decisions
 
-If you don’t have a local GPU or builds/searches are too slow, use one or more of the options below.
+### When to Disable Recomputation
 
-### 1) Use OpenAI embeddings (no local compute)
+LEANN's recomputation feature provides exact distance calculations but can be disabled for extreme QPS requirements:
-
-Fastest path with zero local GPU requirements. Set your API key and use OpenAI embeddings during build and search:
 
 ```bash
-export OPENAI_API_KEY=sk-...
+--no-recompute  # Disable selective recomputation
-
-# Build with OpenAI embeddings
-leann build my-index \
-  --embedding-mode openai \
-  --embedding-model text-embedding-3-small
-
-# Search with OpenAI embeddings (recompute at query time)
-leann search my-index "your query" \
-  --recompute
 ```
 
-### 2) Run remote builds with SkyPilot (cloud GPU)
+**Trade-offs**:
-
+- **With recomputation** (default): Exact distances, best quality, higher latency, minimal storage (only stores metadata, recomputes embeddings on-demand)
-Offload embedding generation and index building to a GPU VM using [SkyPilot](https://skypilot.readthedocs.io/en/latest/). A template is provided at `sky/leann-build.yaml`.
+- **Without recomputation**: Must store full embeddings, significantly higher memory and storage usage (10-100x more), but faster search
-
-```bash
-# One-time: install and configure SkyPilot
-pip install skypilot
-
-# Launch with defaults (L4:1) and mount ./data to ~/leann-data; the build runs automatically
-sky launch -c leann-gpu sky/leann-build.yaml
-
-# Override parameters via -e key=value (optional)
-sky launch -c leann-gpu sky/leann-build.yaml \
-  -e index_name=my-index \
-  -e backend=hnsw \
-  -e embedding_mode=sentence-transformers \
-  -e embedding_model=Qwen/Qwen3-Embedding-0.6B
-
-# Copy the built index back to your local .leann (use rsync)
-rsync -Pavz leann-gpu:~/.leann/indexes/my-index ./.leann/indexes/
-```
-
-### 3) Disable recomputation to trade storage for speed
-
-If you need lower latency and have more storage/memory, disable recomputation. This stores full embeddings and avoids recomputing at search time.
-
-```bash
-# Build without recomputation (HNSW requires non-compact in this mode)
-leann build my-index --no-recompute --no-compact
-
-# Search without recomputation
-leann search my-index "your query" --no-recompute
-```
-
-When to use:
-- Extreme low latency requirements (high QPS, interactive assistants)
-- Read-heavy workloads where storage is cheaper than latency
-- No always-available GPU
-
-Constraints:
-- HNSW: when `--no-recompute` is set, LEANN automatically disables compact mode during build
-- DiskANN: supported; `--no-recompute` skips selective recompute during search
-
-Storage impact:
-- Storing N embeddings of dimension D with float32 requires approximately N × D × 4 bytes
-- Example: 1,000,000 chunks × 768 dims × 4 bytes ≈ 2.86 GB (plus graph/metadata)
-
-Converting an existing index (rebuild required):
-```bash
-# Rebuild in-place (ensure you still have original docs or can regenerate chunks)
-leann build my-index --force --no-recompute --no-compact
-```
-
-Python API usage:
-```python
-from leann import LeannSearcher
-
-searcher = LeannSearcher("/path/to/my-index.leann")
-results = searcher.search("your query", top_k=10, recompute_embeddings=False)
-```
-
-Trade-offs:
-- Lower latency and fewer network hops at query time
-- Significantly higher storage (10–100× vs selective recomputation)
-- Slightly larger memory footprint during build and search
-
-Quick benchmark results (`benchmarks/benchmark_no_recompute.py` with 5k texts, complexity=32):
-
-- HNSW
-
-  ```text
-  recompute=True:  search_time=0.818s, size=1.1MB
-  recompute=False: search_time=0.012s, size=16.6MB
-  ```
-
-- DiskANN
-
-  ```text
-  recompute=True:  search_time=0.041s, size=5.9MB
-  recompute=False: search_time=0.013s, size=24.6MB
-  ```
-
-Conclusion:
-- **HNSW**: `no-recompute` is significantly faster (no embedding recomputation) but requires much more storage (stores all embeddings)
-- **DiskANN**: `no-recompute` uses PQ + partial real distances during traversal (slower but higher accuracy), while `recompute=True` uses pure PQ traversal + final reranking (faster traversal, enables build-time partitioning for smaller storage)
-
 
+**Disable when**:
+- You have abundant storage and memory
+- Need extremely low latency (< 100ms)
+- Running a read-heavy workload where storage cost is acceptable
 
 ## Further Reading
 
 - [Lessons Learned Developing LEANN](https://yichuan-w.github.io/blog/lessons_learned_in_dev_leann/)
 - [LEANN Technical Paper](https://arxiv.org/abs/2506.08276)
 - [DiskANN Original Paper](https://papers.nips.cc/paper/2019/file/09853c7fb1d3f8ee67a61b6bf4a7f8e6-Paper.pdf)
-- [SSD-based Graph Partitioning](https://github.com/SonglinLife/SSD_BASED_PLAN)

packages/leann-backend-diskann/CMakeLists.txt

@@ -0,0 +1,8 @@
+# packages/leann-backend-diskann/CMakeLists.txt (simplified version)
+
+cmake_minimum_required(VERSION 3.20)
+project(leann_backend_diskann_wrapper)
+
+# Tell CMake to directly enter the DiskANN submodule and execute its own CMakeLists.txt
+# DiskANN will handle everything itself, including compiling Python bindings
+add_subdirectory(src/third_party/DiskANN)

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

@@ -1,7 +1 @@
 from . import diskann_backend as diskann_backend
-from . import graph_partition
-
-# Export main classes and functions
-from .graph_partition import GraphPartitioner, partition_graph
-
-__all__ = ["GraphPartitioner", "diskann_backend", "graph_partition", "partition_graph"]

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

@@ -4,7 +4,7 @@ import os
 import struct
 import sys
 from pathlib import Path
-from typing import Any, Literal, Optional
+from typing import Any, Literal
 
 import numpy as np
 import psutil
@@ -22,11 +22,6 @@ logger = logging.getLogger(__name__)
 @contextlib.contextmanager
 def suppress_cpp_output_if_needed():
     """Suppress C++ stdout/stderr based on LEANN_LOG_LEVEL"""
-    # In CI we avoid fiddling with low-level file descriptors to prevent aborts
-    if os.getenv("CI") == "true":
-        yield
-        return
-
     log_level = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
 
     # Only suppress if log level is WARNING or higher (ERROR, CRITICAL)
@@ -142,71 +137,6 @@ class DiskannBuilder(LeannBackendBuilderInterface):
     def __init__(self, **kwargs):
         self.build_params = kwargs
 
-    def _safe_cleanup_after_partition(self, index_dir: Path, index_prefix: str):
-        """
-        Safely cleanup files after partition.
-        In partition mode, C++ doesn't read _disk.index content,
-        so we can delete it if all derived files exist.
-        """
-        disk_index_file = index_dir / f"{index_prefix}_disk.index"
-        beam_search_file = index_dir / f"{index_prefix}_disk_beam_search.index"
-
-        # Required files that C++ partition mode needs
-        # Note: C++ generates these with _disk.index suffix
-        disk_suffix = "_disk.index"
-        required_files = [
-            f"{index_prefix}{disk_suffix}_medoids.bin",  # Critical: assert fails if missing
-            # Note: _centroids.bin is not created in single-shot build - C++ handles this automatically
-            f"{index_prefix}_pq_pivots.bin",  # PQ table
-            f"{index_prefix}_pq_compressed.bin",  # PQ compressed vectors
-        ]
-
-        # Check if all required files exist
-        missing_files = []
-        for filename in required_files:
-            file_path = index_dir / filename
-            if not file_path.exists():
-                missing_files.append(filename)
-
-        if missing_files:
-            logger.warning(
-                f"Cannot safely delete _disk.index - missing required files: {missing_files}"
-            )
-            logger.info("Keeping all original files for safety")
-            return
-
-        # Calculate space savings
-        space_saved = 0
-        files_to_delete = []
-
-        if disk_index_file.exists():
-            space_saved += disk_index_file.stat().st_size
-            files_to_delete.append(disk_index_file)
-
-        if beam_search_file.exists():
-            space_saved += beam_search_file.stat().st_size
-            files_to_delete.append(beam_search_file)
-
-        # Safe to delete!
-        for file_to_delete in files_to_delete:
-            try:
-                os.remove(file_to_delete)
-                logger.info(f"✅ Safely deleted: {file_to_delete.name}")
-            except Exception as e:
-                logger.warning(f"Failed to delete {file_to_delete.name}: {e}")
-
-        if space_saved > 0:
-            space_saved_mb = space_saved / (1024 * 1024)
-            logger.info(f"💾 Space saved: {space_saved_mb:.1f} MB")
-
-            # Show what files are kept
-            logger.info("📁 Kept essential files for partition mode:")
-            for filename in required_files:
-                file_path = index_dir / filename
-                if file_path.exists():
-                    size_mb = file_path.stat().st_size / (1024 * 1024)
-                    logger.info(f"  - {filename} ({size_mb:.1f} MB)")
-
     def build(self, data: np.ndarray, ids: list[str], index_path: str, **kwargs):
         path = Path(index_path)
         index_dir = path.parent
@@ -221,17 +151,6 @@ class DiskannBuilder(LeannBackendBuilderInterface):
         _write_vectors_to_bin(data, index_dir / data_filename)
 
         build_kwargs = {**self.build_params, **kwargs}
-
-        # Extract is_recompute from nested backend_kwargs if needed
-        is_recompute = build_kwargs.get("is_recompute", False)
-        if not is_recompute and "backend_kwargs" in build_kwargs:
-            is_recompute = build_kwargs["backend_kwargs"].get("is_recompute", False)
-
-        # Flatten all backend_kwargs parameters to top level for compatibility
-        if "backend_kwargs" in build_kwargs:
-            nested_params = build_kwargs.pop("backend_kwargs")
-            build_kwargs.update(nested_params)
-
         metric_enum = _get_diskann_metrics().get(
             build_kwargs.get("distance_metric", "mips").lower()
         )
@@ -266,30 +185,6 @@ class DiskannBuilder(LeannBackendBuilderInterface):
                     build_kwargs.get("pq_disk_bytes", 0),
                     "",
                 )
-
-            # Auto-partition if is_recompute is enabled
-            if build_kwargs.get("is_recompute", False):
-                logger.info("is_recompute=True, starting automatic graph partitioning...")
-                from .graph_partition import partition_graph
-
-                # Partition the index using absolute paths
-                # Convert to absolute paths to avoid issues with working directory changes
-                absolute_index_dir = Path(index_dir).resolve()
-                absolute_index_prefix_path = str(absolute_index_dir / index_prefix)
-                disk_graph_path, partition_bin_path = partition_graph(
-                    index_prefix_path=absolute_index_prefix_path,
-                    output_dir=str(absolute_index_dir),
-                    partition_prefix=index_prefix,
-                )
-
-                # Safe cleanup: In partition mode, C++ doesn't read _disk.index content
-                # but still needs the derived files (_medoids.bin, _centroids.bin, etc.)
-                self._safe_cleanup_after_partition(index_dir, index_prefix)
-
-                logger.info("✅ Graph partitioning completed successfully!")
-                logger.info(f"  - Disk graph: {disk_graph_path}")
-                logger.info(f"  - Partition file: {partition_bin_path}")
-
         finally:
             temp_data_file = index_dir / data_filename
             if temp_data_file.exists():
@@ -318,26 +213,7 @@ class DiskannSearcher(BaseSearcher):
 
             # For DiskANN, we need to reinitialize the index when zmq_port changes
             # Store the initialization parameters for later use
-            # Note: C++ load method expects the BASE path (without _disk.index suffix)
+            full_index_prefix = str(self.index_dir / self.index_path.stem)
-            # C++ internally constructs: index_prefix + "_disk.index"
-            index_name = self.index_path.stem  # "simple_test.leann" -> "simple_test"
-            diskann_index_prefix = str(self.index_dir / index_name)  # /path/to/simple_test
-            full_index_prefix = diskann_index_prefix  # /path/to/simple_test (base path)
-
-            # Auto-detect partition files and set partition_prefix
-            partition_graph_file = self.index_dir / f"{index_name}_disk_graph.index"
-            partition_bin_file = self.index_dir / f"{index_name}_partition.bin"
-
-            partition_prefix = ""
-            if partition_graph_file.exists() and partition_bin_file.exists():
-                # C++ expects full path prefix, not just filename
-                partition_prefix = str(self.index_dir / index_name)  # /path/to/simple_test
-                logger.info(
-                    f"✅ Detected partition files, using partition_prefix='{partition_prefix}'"
-                )
-            else:
-                logger.debug("No partition files detected, using standard index files")
-
             self._init_params = {
                 "metric_enum": metric_enum,
                 "full_index_prefix": full_index_prefix,
@@ -345,14 +221,8 @@ class DiskannSearcher(BaseSearcher):
                 "num_nodes_to_cache": kwargs.get("num_nodes_to_cache", 0),
                 "cache_mechanism": 1,
                 "pq_prefix": "",
-                "partition_prefix": partition_prefix,
+                "partition_prefix": "",
             }
-
-            # Log partition configuration for debugging
-            if partition_prefix:
-                logger.info(
-                    f"✅ Detected partition files, using partition_prefix='{partition_prefix}'"
-                )
             self._diskannpy = diskannpy
             self._current_zmq_port = None
             self._index = None
@@ -389,7 +259,7 @@ class DiskannSearcher(BaseSearcher):
         prune_ratio: float = 0.0,
         recompute_embeddings: bool = False,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: Optional[int] = None,
+        zmq_port: int | None = None,
         batch_recompute: bool = False,
         dedup_node_dis: bool = False,
         **kwargs,
@@ -441,14 +311,9 @@ class DiskannSearcher(BaseSearcher):
         else:  # "global"
             use_global_pruning = True
 
-        # Strategy:
+        # Perform search with suppressed C++ output based on log level
-        # - Traversal always uses PQ distances
+        use_deferred_fetch = kwargs.get("USE_DEFERRED_FETCH", True)
-        # - If recompute_embeddings=True, do a single final rerank via deferred fetch
+        recompute_neighors = False
-        #   (fetch embeddings for the final candidate set only)
-        # - Do not recompute neighbor distances along the path
-        use_deferred_fetch = True if recompute_embeddings else False
-        recompute_neighors = False  # Expected typo. For backward compatibility.
-
         with suppress_cpp_output_if_needed():
             labels, distances = self._index.batch_search(
                 query,

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

@@ -10,7 +10,6 @@ import sys
 import threading
 import time
 from pathlib import Path
-from typing import Optional
 
 import numpy as np
 import zmq
@@ -33,7 +32,7 @@ if not logger.handlers:
 
 
 def create_diskann_embedding_server(
-    passages_file: Optional[str] = None,
+    passages_file: str | None = None,
     zmq_port: int = 5555,
     model_name: str = "sentence-transformers/all-mpnet-base-v2",
     embedding_mode: str = "sentence-transformers",
@@ -81,8 +80,7 @@ def create_diskann_embedding_server(
     with open(passages_file) as f:
         meta = json.load(f)
 
-    logger.info(f"Loading PassageManager with metadata_file_path: {passages_file}")
+    passages = PassageManager(meta["passage_sources"])
-    passages = PassageManager(meta["passage_sources"], metadata_file_path=passages_file)
     logger.info(
         f"Loaded PassageManager with {len(passages.global_offset_map)} passages from metadata"
     )
@@ -103,9 +101,8 @@ def create_diskann_embedding_server(
         socket.bind(f"tcp://*:{zmq_port}")
         logger.info(f"DiskANN ZMQ REP server listening on port {zmq_port}")
 
-        socket.setsockopt(zmq.RCVTIMEO, 1000)
+        socket.setsockopt(zmq.RCVTIMEO, 300000)
-        socket.setsockopt(zmq.SNDTIMEO, 1000)
+        socket.setsockopt(zmq.SNDTIMEO, 300000)
-        socket.setsockopt(zmq.LINGER, 0)
 
         while True:
             try:
@@ -222,217 +219,30 @@ def create_diskann_embedding_server(
                 traceback.print_exc()
                 raise
 
-    def zmq_server_thread_with_shutdown(shutdown_event):
+    zmq_thread = threading.Thread(target=zmq_server_thread, daemon=True)
-        """ZMQ server thread that respects shutdown signal.
-
-        This creates its own REP socket, binds to zmq_port, and periodically
-        checks shutdown_event using recv timeouts to exit cleanly.
-        """
-        logger.info("DiskANN ZMQ server thread started with shutdown support")
-
-        context = zmq.Context()
-        rep_socket = context.socket(zmq.REP)
-        rep_socket.bind(f"tcp://*:{zmq_port}")
-        logger.info(f"DiskANN ZMQ REP server listening on port {zmq_port}")
-
-        # Set receive timeout so we can check shutdown_event periodically
-        rep_socket.setsockopt(zmq.RCVTIMEO, 1000)  # 1 second timeout
-        rep_socket.setsockopt(zmq.SNDTIMEO, 1000)
-        rep_socket.setsockopt(zmq.LINGER, 0)
-
-        try:
-            while not shutdown_event.is_set():
-                try:
-                    e2e_start = time.time()
-                    # REP socket receives single-part messages
-                    message = rep_socket.recv()
-
-                    # Check for empty messages - REP socket requires response to every request
-                    if not message:
-                        logger.warning("Received empty message, sending empty response")
-                        rep_socket.send(b"")
-                        continue
-
-                    # Try protobuf first (same logic as original)
-                    texts = []
-                    is_text_request = False
-
-                    try:
-                        req_proto = embedding_pb2.NodeEmbeddingRequest()
-                        req_proto.ParseFromString(message)
-                        node_ids = list(req_proto.node_ids)
-
-                        # Look up texts by node IDs
-                        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:
-                                raise RuntimeError(f"FATAL: Passage with ID {nid} not found")
-
-                        logger.info(f"ZMQ received protobuf request for {len(node_ids)} node IDs")
-                    except Exception:
-                        # Fallback to msgpack for text requests
-                        try:
-                            import msgpack
-
-                            request = msgpack.unpackb(message)
-                            if isinstance(request, list) and all(
-                                isinstance(item, str) for item in request
-                            ):
-                                texts = request
-                                is_text_request = True
-                                logger.info(
-                                    f"ZMQ received msgpack text request for {len(texts)} texts"
-                                )
-                            else:
-                                raise ValueError("Not a valid msgpack text request")
-                        except Exception:
-                            logger.error("Both protobuf and msgpack parsing failed!")
-                            # Send error response
-                            resp_proto = embedding_pb2.NodeEmbeddingResponse()
-                            rep_socket.send(resp_proto.SerializeToString())
-                            continue
-
-                    # Process the request
-                    embeddings = compute_embeddings(texts, model_name, mode=embedding_mode)
-                    logger.info(f"Computed embeddings shape: {embeddings.shape}")
-
-                    # Validation
-                    if np.isnan(embeddings).any() or np.isinf(embeddings).any():
-                        logger.error("NaN or Inf detected in embeddings!")
-                        # Send error response
-                        if is_text_request:
-                            import msgpack
-
-                            response_data = msgpack.packb([])
-                        else:
-                            resp_proto = embedding_pb2.NodeEmbeddingResponse()
-                            response_data = resp_proto.SerializeToString()
-                        rep_socket.send(response_data)
-                        continue
-
-                    # Prepare response based on request type
-                    if is_text_request:
-                        # For direct text requests, return msgpack
-                        import msgpack
-
-                        response_data = msgpack.packb(embeddings.tolist())
-                    else:
-                        # For protobuf requests, return protobuf
-                        resp_proto = embedding_pb2.NodeEmbeddingResponse()
-                        hidden_contiguous = np.ascontiguousarray(embeddings, 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])
-
-                        response_data = resp_proto.SerializeToString()
-
-                    # Send response back to the client
-                    rep_socket.send(response_data)
-
-                    e2e_end = time.time()
-                    logger.info(f"⏱️  ZMQ E2E time: {e2e_end - e2e_start:.6f}s")
-
-                except zmq.Again:
-                    # Timeout - check shutdown_event and continue
-                    continue
-                except Exception as e:
-                    if not shutdown_event.is_set():
-                        logger.error(f"Error in ZMQ server loop: {e}")
-                        try:
-                            # Send error response for REP socket
-                            resp_proto = embedding_pb2.NodeEmbeddingResponse()
-                            rep_socket.send(resp_proto.SerializeToString())
-                        except Exception:
-                            pass
-                    else:
-                        logger.info("Shutdown in progress, ignoring ZMQ error")
-                        break
-        finally:
-            try:
-                rep_socket.close(0)
-            except Exception:
-                pass
-            try:
-                context.term()
-            except Exception:
-                pass
-
-        logger.info("DiskANN ZMQ server thread exiting gracefully")
-
-    # Add shutdown coordination
-    shutdown_event = threading.Event()
-
-    def shutdown_zmq_server():
-        """Gracefully shutdown ZMQ server."""
-        logger.info("Initiating graceful shutdown...")
-        shutdown_event.set()
-
-        if zmq_thread.is_alive():
-            logger.info("Waiting for ZMQ thread to finish...")
-            zmq_thread.join(timeout=5)
-            if zmq_thread.is_alive():
-                logger.warning("ZMQ thread did not finish in time")
-
-        # Clean up ZMQ resources
-        try:
-            # Note: socket and context are cleaned up by thread exit
-            logger.info("ZMQ resources cleaned up")
-        except Exception as e:
-            logger.warning(f"Error cleaning ZMQ resources: {e}")
-
-        # Clean up other resources
-        try:
-            import gc
-
-            gc.collect()
-            logger.info("Additional resources cleaned up")
-        except Exception as e:
-            logger.warning(f"Error cleaning additional resources: {e}")
-
-        logger.info("Graceful shutdown completed")
-        sys.exit(0)
-
-    # Register signal handlers within this function scope
-    import signal
-
-    def signal_handler(sig, frame):
-        logger.info(f"Received signal {sig}, shutting down gracefully...")
-        shutdown_zmq_server()
-
-    signal.signal(signal.SIGTERM, signal_handler)
-    signal.signal(signal.SIGINT, signal_handler)
-
-    # Start ZMQ thread (NOT daemon!)
-    zmq_thread = threading.Thread(
-        target=lambda: zmq_server_thread_with_shutdown(shutdown_event),
-        daemon=False,  # Not daemon - we want to wait for it
-    )
     zmq_thread.start()
     logger.info(f"Started DiskANN ZMQ server thread on port {zmq_port}")
 
     # Keep the main thread alive
     try:
-        while not shutdown_event.is_set():
+        while True:
-            time.sleep(0.1)  # Check shutdown more frequently
+            time.sleep(1)
     except KeyboardInterrupt:
         logger.info("DiskANN Server shutting down...")
-        shutdown_zmq_server()
         return
 
-    # If we reach here, shutdown was triggered by signal
-    logger.info("Main loop exited, process should be shutting down")
-
 
 if __name__ == "__main__":
+    import signal
     import sys
 
-    # Signal handlers are now registered within create_diskann_embedding_server
+    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")
@@ -451,7 +261,7 @@ if __name__ == "__main__":
         "--embedding-mode",
         type=str,
         default="sentence-transformers",
-        choices=["sentence-transformers", "openai", "mlx", "ollama"],
+        choices=["sentence-transformers", "openai", "mlx"],
         help="Embedding backend mode",
     )
     parser.add_argument(

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

@@ -1,299 +0,0 @@
-#!/usr/bin/env python3
-"""
-Graph Partition Module for LEANN DiskANN Backend
-
-This module provides Python bindings for the graph partition functionality
-of DiskANN, allowing users to partition disk-based indices for better
-performance.
-"""
-
-import os
-import shutil
-import subprocess
-import tempfile
-from pathlib import Path
-from typing import Optional
-
-
-class GraphPartitioner:
-    """
-    A Python interface for DiskANN's graph partition functionality.
-
-    This class provides methods to partition disk-based indices for improved
-    search performance and memory efficiency.
-    """
-
-    def __init__(self, build_type: str = "release"):
-        """
-        Initialize the GraphPartitioner.
-
-        Args:
-            build_type: Build type for the executables ("debug" or "release")
-        """
-        self.build_type = build_type
-        self._ensure_executables()
-
-    def _get_executable_path(self, name: str) -> str:
-        """Get the path to a graph partition executable."""
-        # Get the directory where this Python module is located
-        module_dir = Path(__file__).parent
-        # Navigate to the graph_partition directory
-        graph_partition_dir = module_dir.parent / "third_party" / "DiskANN" / "graph_partition"
-        executable_path = graph_partition_dir / "build" / self.build_type / "graph_partition" / name
-
-        if not executable_path.exists():
-            raise FileNotFoundError(f"Executable {name} not found at {executable_path}")
-
-        return str(executable_path)
-
-    def _ensure_executables(self):
-        """Ensure that the required executables are built."""
-        try:
-            self._get_executable_path("partitioner")
-            self._get_executable_path("index_relayout")
-        except FileNotFoundError:
-            # Try to build the executables automatically
-            print("Executables not found, attempting to build them...")
-            self._build_executables()
-
-    def _build_executables(self):
-        """Build the required executables."""
-        graph_partition_dir = (
-            Path(__file__).parent.parent / "third_party" / "DiskANN" / "graph_partition"
-        )
-        original_dir = os.getcwd()
-
-        try:
-            os.chdir(graph_partition_dir)
-
-            # Clean any existing build
-            if (graph_partition_dir / "build").exists():
-                shutil.rmtree(graph_partition_dir / "build")
-
-            # Run the build script
-            cmd = ["./build.sh", self.build_type, "split_graph", "/tmp/dummy"]
-            subprocess.run(cmd, capture_output=True, text=True, cwd=graph_partition_dir)
-
-            # Check if executables were created
-            partitioner_path = self._get_executable_path("partitioner")
-            relayout_path = self._get_executable_path("index_relayout")
-
-            print(f"✅ Built partitioner: {partitioner_path}")
-            print(f"✅ Built index_relayout: {relayout_path}")
-
-        except Exception as e:
-            raise RuntimeError(f"Failed to build executables: {e}")
-        finally:
-            os.chdir(original_dir)
-
-    def partition_graph(
-        self,
-        index_prefix_path: str,
-        output_dir: Optional[str] = None,
-        partition_prefix: Optional[str] = None,
-        **kwargs,
-    ) -> tuple[str, str]:
-        """
-        Partition a disk-based index for improved performance.
-
-        Args:
-            index_prefix_path: Path to the index prefix (e.g., "/path/to/index")
-            output_dir: Output directory for results (defaults to parent of index_prefix_path)
-            partition_prefix: Prefix for output files (defaults to basename of index_prefix_path)
-            **kwargs: Additional parameters for graph partitioning:
-                - gp_times: Number of LDG partition iterations (default: 10)
-                - lock_nums: Number of lock nodes (default: 10)
-                - cut: Cut adjacency list degree (default: 100)
-                - scale_factor: Scale factor (default: 1)
-                - data_type: Data type (default: "float")
-                - thread_nums: Number of threads (default: 10)
-
-        Returns:
-            Tuple of (disk_graph_index_path, partition_bin_path)
-
-        Raises:
-            RuntimeError: If the partitioning process fails
-        """
-        # Set default parameters
-        params = {
-            "gp_times": 10,
-            "lock_nums": 10,
-            "cut": 100,
-            "scale_factor": 1,
-            "data_type": "float",
-            "thread_nums": 10,
-            **kwargs,
-        }
-
-        # Determine output directory
-        if output_dir is None:
-            output_dir = str(Path(index_prefix_path).parent)
-
-        # Create output directory if it doesn't exist
-        Path(output_dir).mkdir(parents=True, exist_ok=True)
-
-        # Determine partition prefix
-        if partition_prefix is None:
-            partition_prefix = Path(index_prefix_path).name
-
-        # Get executable paths
-        partitioner_path = self._get_executable_path("partitioner")
-        relayout_path = self._get_executable_path("index_relayout")
-
-        # Create temporary directory for processing
-        with tempfile.TemporaryDirectory() as temp_dir:
-            # Change to the graph_partition directory for temporary files
-            graph_partition_dir = (
-                Path(__file__).parent.parent / "third_party" / "DiskANN" / "graph_partition"
-            )
-            original_dir = os.getcwd()
-
-            try:
-                os.chdir(graph_partition_dir)
-
-                # Create temporary data directory
-                temp_data_dir = Path(temp_dir) / "data"
-                temp_data_dir.mkdir(parents=True, exist_ok=True)
-
-                # Set up paths for temporary files
-                graph_path = temp_data_dir / "starling" / "_M_R_L_B" / "GRAPH"
-                graph_gp_path = (
-                    graph_path
-                    / f"GP_TIMES_{params['gp_times']}_LOCK_{params['lock_nums']}_GP_USE_FREQ0_CUT{params['cut']}_SCALE{params['scale_factor']}"
-                )
-                graph_gp_path.mkdir(parents=True, exist_ok=True)
-
-                # Find input index file
-                old_index_file = f"{index_prefix_path}_disk_beam_search.index"
-                if not os.path.exists(old_index_file):
-                    old_index_file = f"{index_prefix_path}_disk.index"
-
-                if not os.path.exists(old_index_file):
-                    raise RuntimeError(f"Index file not found: {old_index_file}")
-
-                # Run partitioner
-                gp_file_path = graph_gp_path / "_part.bin"
-                partitioner_cmd = [
-                    partitioner_path,
-                    "--index_file",
-                    old_index_file,
-                    "--data_type",
-                    params["data_type"],
-                    "--gp_file",
-                    str(gp_file_path),
-                    "-T",
-                    str(params["thread_nums"]),
-                    "--ldg_times",
-                    str(params["gp_times"]),
-                    "--scale",
-                    str(params["scale_factor"]),
-                    "--mode",
-                    "1",
-                ]
-
-                print(f"Running partitioner: {' '.join(partitioner_cmd)}")
-                result = subprocess.run(
-                    partitioner_cmd, capture_output=True, text=True, cwd=graph_partition_dir
-                )
-
-                if result.returncode != 0:
-                    raise RuntimeError(
-                        f"Partitioner failed with return code {result.returncode}.\n"
-                        f"stdout: {result.stdout}\n"
-                        f"stderr: {result.stderr}"
-                    )
-
-                # Run relayout
-                part_tmp_index = graph_gp_path / "_part_tmp.index"
-                relayout_cmd = [
-                    relayout_path,
-                    old_index_file,
-                    str(gp_file_path),
-                    params["data_type"],
-                    "1",
-                ]
-
-                print(f"Running relayout: {' '.join(relayout_cmd)}")
-                result = subprocess.run(
-                    relayout_cmd, capture_output=True, text=True, cwd=graph_partition_dir
-                )
-
-                if result.returncode != 0:
-                    raise RuntimeError(
-                        f"Relayout failed with return code {result.returncode}.\n"
-                        f"stdout: {result.stdout}\n"
-                        f"stderr: {result.stderr}"
-                    )
-
-                # Copy results to output directory
-                disk_graph_path = Path(output_dir) / f"{partition_prefix}_disk_graph.index"
-                partition_bin_path = Path(output_dir) / f"{partition_prefix}_partition.bin"
-
-                shutil.copy2(part_tmp_index, disk_graph_path)
-                shutil.copy2(gp_file_path, partition_bin_path)
-
-                print(f"Results copied to: {output_dir}")
-                return str(disk_graph_path), str(partition_bin_path)
-
-            finally:
-                os.chdir(original_dir)
-
-    def get_partition_info(self, partition_bin_path: str) -> dict:
-        """
-        Get information about a partition file.
-
-        Args:
-            partition_bin_path: Path to the partition binary file
-
-        Returns:
-            Dictionary containing partition information
-        """
-        if not os.path.exists(partition_bin_path):
-            raise FileNotFoundError(f"Partition file not found: {partition_bin_path}")
-
-        # For now, return basic file information
-        # In the future, this could parse the binary file for detailed info
-        stat = os.stat(partition_bin_path)
-        return {
-            "file_size": stat.st_size,
-            "file_path": partition_bin_path,
-            "modified_time": stat.st_mtime,
-        }
-
-
-def partition_graph(
-    index_prefix_path: str,
-    output_dir: Optional[str] = None,
-    partition_prefix: Optional[str] = None,
-    build_type: str = "release",
-    **kwargs,
-) -> tuple[str, str]:
-    """
-    Convenience function to partition a graph index.
-
-    Args:
-        index_prefix_path: Path to the index prefix
-        output_dir: Output directory (defaults to parent of index_prefix_path)
-        partition_prefix: Prefix for output files (defaults to basename of index_prefix_path)
-        build_type: Build type for executables ("debug" or "release")
-        **kwargs: Additional parameters for graph partitioning
-
-    Returns:
-        Tuple of (disk_graph_index_path, partition_bin_path)
-    """
-    partitioner = GraphPartitioner(build_type=build_type)
-    return partitioner.partition_graph(index_prefix_path, output_dir, partition_prefix, **kwargs)
-
-
-# Example usage:
-if __name__ == "__main__":
-    # Example: partition an index
-    try:
-        disk_graph_path, partition_bin_path = partition_graph(
-            "/path/to/your/index_prefix", gp_times=10, lock_nums=10, cut=100
-        )
-        print("Partitioning completed successfully!")
-        print(f"Disk graph index: {disk_graph_path}")
-        print(f"Partition binary: {partition_bin_path}")
-    except Exception as e:
-        print(f"Partitioning failed: {e}")

packages/leann-backend-diskann/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "leann-backend-diskann"
-version = "0.2.9"
+version = "0.2.1"
-dependencies = ["leann-core==0.2.9", "numpy", "protobuf>=3.19.0"]
+dependencies = ["leann-core==0.2.1", "numpy", "protobuf>=3.19.0"]
 
 [tool.scikit-build]
 # Key: simplified CMake path
@@ -17,5 +17,3 @@ editable.mode = "redirect"
 cmake.build-type = "Release"
 build.verbose = true
 build.tool-args = ["-j8"]
-# Let CMake find packages via Homebrew prefix
-cmake.define = {CMAKE_PREFIX_PATH = {env = "CMAKE_PREFIX_PATH"}, OpenMP_ROOT = {env = "OpenMP_ROOT"}}

packages/leann-backend-hnsw/CMakeLists.txt

@@ -5,20 +5,11 @@ set(CMAKE_CXX_COMPILER_WORKS 1)
 
 # Set OpenMP path for macOS
 if(APPLE)
-    # Detect Homebrew installation path (Apple Silicon vs Intel)
+    set(OpenMP_C_FLAGS "-Xpreprocessor -fopenmp -I/opt/homebrew/opt/libomp/include")
-    if(EXISTS "/opt/homebrew/opt/libomp")
+    set(OpenMP_CXX_FLAGS "-Xpreprocessor -fopenmp -I/opt/homebrew/opt/libomp/include")
-        set(HOMEBREW_PREFIX "/opt/homebrew")
-    elseif(EXISTS "/usr/local/opt/libomp")
-        set(HOMEBREW_PREFIX "/usr/local")
-    else()
-        message(FATAL_ERROR "Could not find libomp installation. Please install with: brew install libomp")
-    endif()
-
-    set(OpenMP_C_FLAGS "-Xpreprocessor -fopenmp -I${HOMEBREW_PREFIX}/opt/libomp/include")
-    set(OpenMP_CXX_FLAGS "-Xpreprocessor -fopenmp -I${HOMEBREW_PREFIX}/opt/libomp/include")
     set(OpenMP_C_LIB_NAMES "omp")
     set(OpenMP_CXX_LIB_NAMES "omp")
-    set(OpenMP_omp_LIBRARY "${HOMEBREW_PREFIX}/opt/libomp/lib/libomp.dylib")
+    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++")

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

@@ -1,6 +1,5 @@
 import argparse
 import gc  # Import garbage collector interface
-import logging
 import os
 import struct
 import sys
@@ -8,12 +7,6 @@ import time
 
 import numpy as np
 
-# Set up logging to avoid print buffer issues
-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)
-
 # --- FourCCs (add more if needed) ---
 INDEX_HNSW_FLAT_FOURCC = int.from_bytes(b"IHNf", "little")
 # Add other HNSW fourccs if you expect different storage types inside HNSW
@@ -250,8 +243,6 @@ def convert_hnsw_graph_to_csr(input_filename, output_filename, prune_embeddings=
         output_filename: Output CSR index file
         prune_embeddings: Whether to prune embedding storage (write NULL storage marker)
     """
-    # Keep prints simple; rely on CI runner to flush output as needed
-
     print(f"Starting conversion: {input_filename} -> {output_filename}")
     start_time = time.time()
     original_hnsw_data = {}

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

@@ -2,7 +2,7 @@ import logging
 import os
 import shutil
 from pathlib import Path
-from typing import Any, Literal, Optional
+from typing import Any, Literal
 
 import numpy as np
 from leann.interface import (
@@ -54,13 +54,12 @@ 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 and self.is_compact:
+        if not self.is_recompute:
-            # Auto-correct: non-recompute requires non-compact storage for HNSW
+            if self.is_compact:
-            logger.warning(
+                # TODO: support this case @andy
-                "is_recompute=False requires non-compact HNSW. Forcing is_compact=False."
+                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."
-            self.is_compact = False
+                )
-            self.build_params["is_compact"] = False
 
     def build(self, data: np.ndarray, ids: list[str], index_path: str, **kwargs):
         from . import faiss  # type: ignore
@@ -153,7 +152,7 @@ class HNSWSearcher(BaseSearcher):
         self,
         query: np.ndarray,
         top_k: int,
-        zmq_port: Optional[int] = None,
+        zmq_port: int | None = None,
         complexity: int = 64,
         beam_width: int = 1,
         prune_ratio: float = 0.0,
@@ -185,11 +184,9 @@ class HNSWSearcher(BaseSearcher):
         """
         from . import faiss  # type: ignore
 
-        if not recompute_embeddings and self.is_pruned:
+        if not recompute_embeddings:
-            raise RuntimeError(
+            if self.is_pruned:
-                "Recompute is required for pruned/compact HNSW index. "
+                raise RuntimeError("Recompute is required for pruned index.")
-                "Re-run search with --recompute, or rebuild with --no-recompute and --no-compact."
-            )
         if recompute_embeddings:
             if zmq_port is None:
                 raise ValueError("zmq_port must be provided if recompute_embeddings is True")

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

@@ -10,7 +10,6 @@ import sys
 import threading
 import time
 from pathlib import Path
-from typing import Optional
 
 import msgpack
 import numpy as np
@@ -34,7 +33,7 @@ if not logger.handlers:
 
 
 def create_hnsw_embedding_server(
-    passages_file: Optional[str] = None,
+    passages_file: str | None = None,
     zmq_port: int = 5555,
     model_name: str = "sentence-transformers/all-mpnet-base-v2",
     distance_metric: str = "mips",
@@ -82,317 +81,199 @@ def create_hnsw_embedding_server(
     with open(passages_file) as f:
         meta = json.load(f)
 
-    # Let PassageManager handle path resolution uniformly. It supports fallback order:
+    # Convert relative paths to absolute paths based on metadata file location
-    # 1) path/index_path; 2) *_relative; 3) standard siblings next to meta
+    metadata_dir = Path(passages_file).parent.parent  # Go up one level from the metadata file
-    passages = PassageManager(meta["passage_sources"], metadata_file_path=passages_file)
+    passage_sources = []
-    # Dimension from metadata for shaping responses
+    for source in meta["passage_sources"]:
-    try:
+        source_copy = source.copy()
-        embedding_dim: int = int(meta.get("dimensions", 0))
+        # Convert relative paths to absolute paths
-    except Exception:
+        if not Path(source_copy["path"]).is_absolute():
-        embedding_dim = 0
+            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"
     )
 
-    # (legacy ZMQ thread removed; using shutdown-capable server only)
+    def zmq_server_thread():
-
+        """ZMQ server thread"""
-    def zmq_server_thread_with_shutdown(shutdown_event):
-        """ZMQ server thread that respects shutdown signal.
-
-        Creates its own REP socket bound to zmq_port and polls with timeouts
-        to allow graceful shutdown.
-        """
-        logger.info("ZMQ server thread started with shutdown support")
-
         context = zmq.Context()
-        rep_socket = context.socket(zmq.REP)
+        socket = context.socket(zmq.REP)
-        rep_socket.bind(f"tcp://*:{zmq_port}")
+        socket.bind(f"tcp://*:{zmq_port}")
-        logger.info(f"HNSW ZMQ REP server listening on port {zmq_port}")
+        logger.info(f"HNSW ZMQ server listening on port {zmq_port}")
-        rep_socket.setsockopt(zmq.RCVTIMEO, 1000)
-        # Keep sends from blocking during shutdown; fail fast and drop on close
-        rep_socket.setsockopt(zmq.SNDTIMEO, 1000)
-        rep_socket.setsockopt(zmq.LINGER, 0)
 
-        # Track last request type/length for shape-correct fallbacks
+        socket.setsockopt(zmq.RCVTIMEO, 300000)
-        last_request_type = "unknown"  # 'text' | 'distance' | 'embedding' | 'unknown'
+        socket.setsockopt(zmq.SNDTIMEO, 300000)
-        last_request_length = 0
 
-        try:
+        while True:
-            while not shutdown_event.is_set():
+            try:
-                try:
+                message_bytes = socket.recv()
-                    e2e_start = time.time()
+                logger.debug(f"Received ZMQ request of size {len(message_bytes)} bytes")
-                    logger.debug("🔍 Waiting for ZMQ message...")
-                    request_bytes = rep_socket.recv()
 
-                    # Rest of the processing logic (same as original)
+                e2e_start = time.time()
-                    request = msgpack.unpackb(request_bytes)
+                request_payload = msgpack.unpackb(message_bytes)
 
-                    if len(request) == 1 and request[0] == "__QUERY_MODEL__":
+                # Handle direct text embedding request
-                        response_bytes = msgpack.packb([model_name])
+                if isinstance(request_payload, list) and len(request_payload) > 0:
-                        rep_socket.send(response_bytes)
+                    # Check if this is a direct text request (list of strings)
-                        continue
+                    if all(isinstance(item, str) for item in request_payload):
+                        logger.info(
+                            f"Processing direct text embedding request for {len(request_payload)} texts in {embedding_mode} mode"
+                        )
 
-                    # Handle direct text embedding request
+                        # Use unified embedding computation (now with model caching)
-                    if (
+                        embeddings = compute_embeddings(
-                        isinstance(request, list)
+                            request_payload, model_name, mode=embedding_mode
-                        and request
+                        )
-                        and all(isinstance(item, str) for item in request)
+
-                    ):
+                        response = embeddings.tolist()
-                        last_request_type = "text"
+                        socket.send(msgpack.packb(response))
-                        last_request_length = len(request)
-                        embeddings = compute_embeddings(request, model_name, mode=embedding_mode)
-                        rep_socket.send(msgpack.packb(embeddings.tolist()))
                         e2e_end = time.time()
                         logger.info(f"⏱️  Text embedding E2E time: {e2e_end - e2e_start:.6f}s")
                         continue
 
-                    # Handle distance calculation request: [[ids], [query_vector]]
+                # Handle distance calculation requests
-                    if (
+                if (
-                        isinstance(request, list)
+                    isinstance(request_payload, list)
-                        and len(request) == 2
+                    and len(request_payload) == 2
-                        and isinstance(request[0], list)
+                    and isinstance(request_payload[0], list)
-                        and isinstance(request[1], list)
+                    and isinstance(request_payload[1], list)
-                    ):
+                ):
-                        node_ids = request[0]
+                    node_ids = request_payload[0]
-                        # Handle nested [[ids]] shape defensively
+                    query_vector = np.array(request_payload[1], dtype=np.float32)
-                        if len(node_ids) == 1 and isinstance(node_ids[0], list):
-                            node_ids = node_ids[0]
-                        query_vector = np.array(request[1], dtype=np.float32)
-                        last_request_type = "distance"
-                        last_request_length = len(node_ids)
 
-                        logger.debug("Distance calculation request received")
+                    logger.debug("Distance calculation request received")
-                        logger.debug(f"    Node IDs: {node_ids}")
+                    logger.debug(f"    Node IDs: {node_ids}")
-                        logger.debug(f"    Query vector dim: {len(query_vector)}")
+                    logger.debug(f"    Query vector dim: {len(query_vector)}")
 
-                        # Gather texts for found ids
+                    # Get embeddings for node IDs
-                        texts: list[str] = []
+                    texts = []
-                        found_indices: list[int] = []
+                    for nid in node_ids:
-                        for idx, nid in enumerate(node_ids):
-                            try:
-                                passage_data = passages.get_passage(str(nid))
-                                txt = passage_data.get("text", "")
-                                if isinstance(txt, str) and len(txt) > 0:
-                                    texts.append(txt)
-                                    found_indices.append(idx)
-                                else:
-                                    logger.error(f"Empty text for passage ID {nid}")
-                            except KeyError:
-                                logger.error(f"Passage ID {nid} not found")
-                            except Exception as e:
-                                logger.error(f"Exception looking up passage ID {nid}: {e}")
-
-                        # Prepare full-length response with large sentinel values
-                        large_distance = 1e9
-                        response_distances = [large_distance] * len(node_ids)
-
-                        if texts:
-                            try:
-                                embeddings = compute_embeddings(
-                                    texts, model_name, mode=embedding_mode
-                                )
-                                logger.info(
-                                    f"Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
-                                )
-                                if distance_metric == "l2":
-                                    partial = np.sum(
-                                        np.square(embeddings - query_vector.reshape(1, -1)), axis=1
-                                    )
-                                else:  # mips or cosine
-                                    partial = -np.dot(embeddings, query_vector)
-
-                                for pos, dval in zip(found_indices, partial.flatten().tolist()):
-                                    response_distances[pos] = float(dval)
-                            except Exception as e:
-                                logger.error(f"Distance computation error, using sentinels: {e}")
-
-                        # Send response in expected shape [[distances]]
-                        rep_socket.send(msgpack.packb([response_distances], use_single_float=True))
-                        e2e_end = time.time()
-                        logger.info(f"⏱️  Distance calculation E2E time: {e2e_end - e2e_start:.6f}s")
-                        continue
-
-                    # Fallback: treat as embedding-by-id request
-                    if (
-                        isinstance(request, list)
-                        and len(request) == 1
-                        and isinstance(request[0], list)
-                    ):
-                        node_ids = request[0]
-                    elif isinstance(request, list):
-                        node_ids = request
-                    else:
-                        node_ids = []
-                    last_request_type = "embedding"
-                    last_request_length = len(node_ids)
-                    logger.info(f"ZMQ received {len(node_ids)} node IDs for embedding fetch")
-
-                    # Preallocate zero-filled flat data for robustness
-                    if embedding_dim <= 0:
-                        dims = [0, 0]
-                        flat_data: list[float] = []
-                    else:
-                        dims = [len(node_ids), embedding_dim]
-                        flat_data = [0.0] * (dims[0] * dims[1])
-
-                    # Collect texts for found ids
-                    texts: list[str] = []
-                    found_indices: list[int] = []
-                    for idx, nid in enumerate(node_ids):
                         try:
                             passage_data = passages.get_passage(str(nid))
-                            txt = passage_data.get("text", "")
+                            txt = passage_data["text"]
-                            if isinstance(txt, str) and len(txt) > 0:
+                            texts.append(txt)
-                                texts.append(txt)
-                                found_indices.append(idx)
-                            else:
-                                logger.error(f"Empty text for passage ID {nid}")
                         except KeyError:
-                            logger.error(f"Passage with 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:
                             logger.error(f"Exception looking up passage ID {nid}: {e}")
+                            raise
 
-                    if texts:
+                    # Process embeddings
-                        try:
+                    embeddings = compute_embeddings(texts, model_name, mode=embedding_mode)
-                            embeddings = compute_embeddings(texts, model_name, mode=embedding_mode)
+                    logger.info(
-                            logger.info(
+                        f"Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
-                                f"Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
+                    )
-                            )
 
-                            if np.isnan(embeddings).any() or np.isinf(embeddings).any():
+                    # Calculate distances
-                                logger.error(
+                    if distance_metric == "l2":
-                                    f"NaN or Inf detected in embeddings! Requested IDs: {node_ids[:5]}..."
+                        distances = np.sum(
-                                )
+                            np.square(embeddings - query_vector.reshape(1, -1)), axis=1
-                                dims = [0, embedding_dim]
+                        )
-                                flat_data = []
+                    else:  # mips or cosine
-                            else:
+                        distances = -np.dot(embeddings, query_vector)
-                                emb_f32 = np.ascontiguousarray(embeddings, dtype=np.float32)
-                                flat = emb_f32.flatten().tolist()
-                                for j, pos in enumerate(found_indices):
-                                    start = pos * embedding_dim
-                                    end = start + embedding_dim
-                                    if end <= len(flat_data):
-                                        flat_data[start:end] = flat[
-                                            j * embedding_dim : (j + 1) * embedding_dim
-                                        ]
-                        except Exception as e:
-                            logger.error(f"Embedding computation error, returning zeros: {e}")
 
-                    response_payload = [dims, flat_data]
+                    response_payload = distances.flatten().tolist()
-                    response_bytes = msgpack.packb(response_payload, use_single_float=True)
+                    response_bytes = msgpack.packb([response_payload], use_single_float=True)
+                    logger.debug(f"Sending distance response with {len(distances)} distances")
 
-                    rep_socket.send(response_bytes)
+                    socket.send(response_bytes)
                     e2e_end = time.time()
-                    logger.info(f"⏱️  ZMQ E2E time: {e2e_end - e2e_start:.6f}s")
+                    logger.info(f"⏱️  Distance calculation E2E time: {e2e_end - e2e_start:.6f}s")
-
-                except zmq.Again:
-                    # Timeout - check shutdown_event and continue
                     continue
-                except Exception as e:
-                    if not shutdown_event.is_set():
-                        logger.error(f"Error in ZMQ server loop: {e}")
-                        # Shape-correct fallback
-                        try:
-                            if last_request_type == "distance":
-                                large_distance = 1e9
-                                fallback_len = max(0, int(last_request_length))
-                                safe = [[large_distance] * fallback_len]
-                            elif last_request_type == "embedding":
-                                bsz = max(0, int(last_request_length))
-                                dim = max(0, int(embedding_dim))
-                                safe = (
-                                    [[bsz, dim], [0.0] * (bsz * dim)] if dim > 0 else [[0, 0], []]
-                                )
-                            elif last_request_type == "text":
-                                safe = []  # direct text embeddings expectation is a flat list
-                            else:
-                                safe = [[0, int(embedding_dim) if embedding_dim > 0 else 0], []]
-                            rep_socket.send(msgpack.packb(safe, use_single_float=True))
-                        except Exception:
-                            pass
-                    else:
-                        logger.info("Shutdown in progress, ignoring ZMQ error")
-                        break
-        finally:
-            try:
-                rep_socket.close(0)
-            except Exception:
-                pass
-            try:
-                context.term()
-            except Exception:
-                pass
 
-        logger.info("ZMQ server thread exiting gracefully")
+                # Standard embedding request (passage ID lookup)
+                if (
+                    not isinstance(request_payload, list)
+                    or len(request_payload) != 1
+                    or not isinstance(request_payload[0], list)
+                ):
+                    logger.error(
+                        f"Invalid MessagePack request format. Expected [[ids...]] or [texts...], got: {type(request_payload)}"
+                    )
+                    socket.send(msgpack.packb([[], []]))
+                    continue
 
-    # Add shutdown coordination
+                node_ids = request_payload[0]
-    shutdown_event = threading.Event()
+                logger.debug(f"Request for {len(node_ids)} node embeddings")
 
-    def shutdown_zmq_server():
+                # Look up texts by node IDs
-        """Gracefully shutdown ZMQ server."""
+                texts = []
-        logger.info("Initiating graceful shutdown...")
+                for nid in node_ids:
-        shutdown_event.set()
+                    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:
+                        raise RuntimeError(f"FATAL: Passage with ID {nid} not found")
+                    except Exception as e:
+                        logger.error(f"Exception looking up passage ID {nid}: {e}")
+                        raise
 
-        if zmq_thread.is_alive():
+                # Process embeddings
-            logger.info("Waiting for ZMQ thread to finish...")
+                embeddings = compute_embeddings(texts, model_name, mode=embedding_mode)
-            zmq_thread.join(timeout=5)
+                logger.info(
-            if zmq_thread.is_alive():
+                    f"Computed embeddings for {len(texts)} texts, shape: {embeddings.shape}"
-                logger.warning("ZMQ thread did not finish in time")
+                )
 
-        # Clean up ZMQ resources
+                # Serialization and response
-        try:
+                if np.isnan(embeddings).any() or np.isinf(embeddings).any():
-            # Note: socket and context are cleaned up by thread exit
+                    logger.error(
-            logger.info("ZMQ resources cleaned up")
+                        f"NaN or Inf detected in embeddings! Requested IDs: {node_ids[:5]}..."
-        except Exception as e:
+                    )
-            logger.warning(f"Error cleaning ZMQ resources: {e}")
+                    raise AssertionError()
 
-        # Clean up other resources
+                hidden_contiguous_f32 = np.ascontiguousarray(embeddings, dtype=np.float32)
-        try:
+                response_payload = [
-            import gc
+                    list(hidden_contiguous_f32.shape),
+                    hidden_contiguous_f32.flatten().tolist(),
+                ]
+                response_bytes = msgpack.packb(response_payload, use_single_float=True)
 
-            gc.collect()
+                socket.send(response_bytes)
-            logger.info("Additional resources cleaned up")
+                e2e_end = time.time()
-        except Exception as e:
+                logger.info(f"⏱️  ZMQ E2E time: {e2e_end - e2e_start:.6f}s")
-            logger.warning(f"Error cleaning additional resources: {e}")
 
-        logger.info("Graceful shutdown completed")
+            except zmq.Again:
-        sys.exit(0)
+                logger.debug("ZMQ socket timeout, continuing to listen")
+                continue
+            except Exception as e:
+                logger.error(f"Error in ZMQ server loop: {e}")
+                import traceback
 
-    # Register signal handlers within this function scope
+                traceback.print_exc()
-    import signal
+                socket.send(msgpack.packb([[], []]))
 
-    def signal_handler(sig, frame):
+    zmq_thread = threading.Thread(target=zmq_server_thread, daemon=True)
-        logger.info(f"Received signal {sig}, shutting down gracefully...")
-        shutdown_zmq_server()
-
-    signal.signal(signal.SIGTERM, signal_handler)
-    signal.signal(signal.SIGINT, signal_handler)
-
-    # Pass shutdown_event to ZMQ thread
-    zmq_thread = threading.Thread(
-        target=lambda: zmq_server_thread_with_shutdown(shutdown_event),
-        daemon=False,  # Not daemon - we want to wait for it
-    )
     zmq_thread.start()
     logger.info(f"Started HNSW ZMQ server thread on port {zmq_port}")
 
     # Keep the main thread alive
     try:
-        while not shutdown_event.is_set():
+        while True:
-            time.sleep(0.1)  # Check shutdown more frequently
+            time.sleep(1)
     except KeyboardInterrupt:
         logger.info("HNSW Server shutting down...")
-        shutdown_zmq_server()
         return
 
-    # If we reach here, shutdown was triggered by signal
-    logger.info("Main loop exited, process should be shutting down")
-
 
 if __name__ == "__main__":
+    import signal
     import sys
 
-    # Signal handlers are now registered within create_hnsw_embedding_server
+    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")
@@ -414,7 +295,7 @@ if __name__ == "__main__":
         "--embedding-mode",
         type=str,
         default="sentence-transformers",
-        choices=["sentence-transformers", "openai", "mlx", "ollama"],
+        choices=["sentence-transformers", "openai", "mlx"],
         help="Embedding backend mode",
     )
 

packages/leann-backend-hnsw/pyproject.toml

@@ -6,10 +6,10 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "leann-backend-hnsw"
-version = "0.2.9"
+version = "0.2.1"
 description = "Custom-built HNSW (Faiss) backend for the Leann toolkit."
 dependencies = [
-    "leann-core==0.2.9",
+    "leann-core==0.2.1",
     "numpy",
     "pyzmq>=23.0.0",
     "msgpack>=1.0.0",
@@ -22,8 +22,6 @@ cmake.build-type = "Release"
 build.verbose = true
 build.tool-args = ["-j8"]
 
-# CMake definitions to optimize compilation and find Homebrew packages
+# CMake definitions to optimize compilation
 [tool.scikit-build.cmake.define]
 CMAKE_BUILD_PARALLEL_LEVEL = "8"
-CMAKE_PREFIX_PATH = {env = "CMAKE_PREFIX_PATH"}
-OpenMP_ROOT = {env = "OpenMP_ROOT"}

packages/leann-core/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "leann-core"
-version = "0.2.9"
+version = "0.2.1"
 description = "Core API and plugin system for LEANN"
 readme = "README.md"
 requires-python = ">=3.9"
@@ -31,10 +31,8 @@ dependencies = [
     "PyPDF2>=3.0.0",
     "pymupdf>=1.23.0",
     "pdfplumber>=0.10.0",
-    "nbconvert>=7.0.0",  # For .ipynb file support
+    "mlx>=0.26.3; sys_platform == 'darwin'",
-    "gitignore-parser>=0.1.12",  # For proper .gitignore handling
+    "mlx-lm>=0.26.0; sys_platform == 'darwin'",
-    "mlx>=0.26.3; sys_platform == 'darwin' and platform_machine == 'arm64'",
-    "mlx-lm>=0.26.0; sys_platform == 'darwin' and platform_machine == 'arm64'",
 ]
 
 [project.optional-dependencies]
@@ -46,7 +44,6 @@ colab = [
 
 [project.scripts]
 leann = "leann.cli:main"
-leann_mcp = "leann.mcp:main"
 
 [tool.setuptools.packages.find]
 where = ["src"]

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

@@ -10,7 +10,7 @@ import time
 import warnings
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any, Literal, Optional
+from typing import Any, Literal
 
 import numpy as np
 
@@ -33,7 +33,7 @@ def compute_embeddings(
     model_name: str,
     mode: str = "sentence-transformers",
     use_server: bool = True,
-    port: Optional[int] = None,
+    port: int | None = None,
     is_build=False,
 ) -> np.ndarray:
     """
@@ -115,62 +115,20 @@ class SearchResult:
 
 
 class PassageManager:
-    def __init__(
+    def __init__(self, passage_sources: list[dict[str, Any]]):
-        self, passage_sources: list[dict[str, Any]], metadata_file_path: Optional[str] = None
-    ):
         self.offset_maps = {}
         self.passage_files = {}
         self.global_offset_map = {}  # Combined map for fast lookup
 
-        # Derive index base name for standard sibling fallbacks, e.g., <index_name>.passages.*
-        index_name_base = None
-        if metadata_file_path:
-            meta_name = Path(metadata_file_path).name
-            if meta_name.endswith(".meta.json"):
-                index_name_base = meta_name[: -len(".meta.json")]
-
         for source in passage_sources:
             assert source["type"] == "jsonl", "only jsonl is supported"
-            passage_file = source.get("path", "")
+            passage_file = source["path"]
-            index_file = source.get("index_path", "")  # .idx file
+            index_file = source["index_path"]  # .idx file
 
-            # Fix path resolution - relative paths should be relative to metadata file directory
+            # Fix path resolution for Colab and other environments
-            def _resolve_candidates(
+            if not Path(index_file).is_absolute():
-                primary: str,
+                # If relative path, try to resolve it properly
-                relative_key: str,
+                index_file = str(Path(index_file).resolve())
-                default_name: Optional[str],
-                source_dict: dict[str, Any],
-            ) -> list[Path]:
-                candidates: list[Path] = []
-                # 1) Primary as-is (absolute or relative)
-                if primary:
-                    p = Path(primary)
-                    candidates.append(p if p.is_absolute() else (Path.cwd() / p))
-                # 2) metadata-relative explicit relative key
-                if metadata_file_path and source_dict.get(relative_key):
-                    candidates.append(Path(metadata_file_path).parent / source_dict[relative_key])
-                # 3) metadata-relative standard sibling filename
-                if metadata_file_path and default_name:
-                    candidates.append(Path(metadata_file_path).parent / default_name)
-                return candidates
-
-            # Build candidate lists and pick first existing; otherwise keep last candidate for error message
-            idx_default = f"{index_name_base}.passages.idx" if index_name_base else None
-            idx_candidates = _resolve_candidates(
-                index_file, "index_path_relative", idx_default, source
-            )
-            pas_default = f"{index_name_base}.passages.jsonl" if index_name_base else None
-            pas_candidates = _resolve_candidates(passage_file, "path_relative", pas_default, source)
-
-            def _pick_existing(cands: list[Path]) -> str:
-                for c in cands:
-                    if c.exists():
-                        return str(c.resolve())
-                # Fallback to last candidate (best guess) even if not exists; will error below
-                return str(cands[-1].resolve()) if cands else ""
-
-            index_file = _pick_existing(idx_candidates)
-            passage_file = _pick_existing(pas_candidates)
 
             if not Path(index_file).exists():
                 raise FileNotFoundError(f"Passage index file not found: {index_file}")
@@ -199,24 +157,12 @@ class LeannBuilder:
         self,
         backend_name: str,
         embedding_model: str = "facebook/contriever",
-        dimensions: Optional[int] = None,
+        dimensions: int | None = None,
         embedding_mode: str = "sentence-transformers",
         **backend_kwargs,
     ):
         self.backend_name = backend_name
-        # Normalize incompatible combinations early (for consistent metadata)
+        backend_factory: LeannBackendFactoryInterface | None = BACKEND_REGISTRY.get(backend_name)
-        if backend_name == "hnsw":
-            is_recompute = backend_kwargs.get("is_recompute", True)
-            is_compact = backend_kwargs.get("is_compact", True)
-            if is_recompute is False and is_compact is True:
-                warnings.warn(
-                    "HNSW with is_recompute=False requires non-compact storage. Forcing is_compact=False.",
-                    UserWarning,
-                    stacklevel=2,
-                )
-                backend_kwargs["is_compact"] = False
-
-        backend_factory: Optional[LeannBackendFactoryInterface] = 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
@@ -296,7 +242,7 @@ class LeannBuilder:
         self.backend_kwargs = backend_kwargs
         self.chunks: list[dict[str, Any]] = []
 
-    def add_text(self, text: str, metadata: Optional[dict[str, Any]] = None):
+    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)))
@@ -368,12 +314,8 @@ class LeannBuilder:
             "passage_sources": [
                 {
                     "type": "jsonl",
-                    # Preserve existing relative file names (backward-compatible)
+                    "path": str(passages_file),
-                    "path": passages_file.name,
+                    "index_path": str(offset_file),
-                    "index_path": offset_file.name,
-                    # Add optional redundant relative keys for remote build portability (non-breaking)
-                    "path_relative": passages_file.name,
-                    "index_path_relative": offset_file.name,
                 }
             ],
         }
@@ -488,12 +430,8 @@ class LeannBuilder:
             "passage_sources": [
                 {
                     "type": "jsonl",
-                    # Preserve existing relative file names (backward-compatible)
+                    "path": str(passages_file),
-                    "path": passages_file.name,
+                    "index_path": str(offset_file),
-                    "index_path": offset_file.name,
-                    # Add optional redundant relative keys for remote build portability (non-breaking)
-                    "path_relative": passages_file.name,
-                    "index_path_relative": offset_file.name,
                 }
             ],
             "built_from_precomputed_embeddings": True,
@@ -535,10 +473,7 @@ class LeannSearcher:
         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")
-        # Delegate portability handling to PassageManager
+        self.passage_manager = PassageManager(self.meta_data.get("passage_sources", []))
-        self.passage_manager = PassageManager(
-            self.meta_data.get("passage_sources", []), metadata_file_path=self.meta_path_str
-        )
         backend_factory = BACKEND_REGISTRY.get(backend_name)
         if backend_factory is None:
             raise ValueError(f"Backend '{backend_name}' not found.")
@@ -611,15 +546,15 @@ class LeannSearcher:
             zmq_port=zmq_port,
             **kwargs,
         )
+        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:
             logger.info(f"  Processing {len(results['labels'][0])} passage IDs:")
-            # Python 3.9 does not support zip(strict=...); lengths are expected to match
             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)
@@ -645,49 +580,19 @@ class LeannSearcher:
                     )
                 except KeyError:
                     RED = "\033[91m"
-                    RESET = "\033[0m"
                     logger.error(
                         f"   {RED}✗{RESET} [{i + 1:2d}] ID: '{string_id}' -> {RED}ERROR: Passage not found!{RESET}"
                     )
 
-        # Define color codes outside the loop for final message
-        GREEN = "\033[92m"
-        RESET = "\033[0m"
         logger.info(f"  {GREEN}✓ Final enriched results: {len(enriched_results)} passages{RESET}")
         return enriched_results
 
-    def cleanup(self):
-        """Explicitly cleanup embedding server resources.
-
-        This method should be called after you're done using the searcher,
-        especially in test environments or batch processing scenarios.
-        """
-        if hasattr(self.backend_impl, "embedding_server_manager"):
-            self.backend_impl.embedding_server_manager.stop_server()
-
-    # Enable automatic cleanup patterns
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc, tb):
-        try:
-            self.cleanup()
-        except Exception:
-            pass
-
-    def __del__(self):
-        try:
-            self.cleanup()
-        except Exception:
-            # Avoid noisy errors during interpreter shutdown
-            pass
-
 
 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,
     ):
@@ -703,7 +608,7 @@ class LeannChat:
         prune_ratio: float = 0.0,
         recompute_embeddings: bool = True,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        llm_kwargs: Optional[dict[str, Any]] = None,
+        llm_kwargs: dict[str, Any] | None = None,
         expected_zmq_port: int = 5557,
         **search_kwargs,
     ):
@@ -751,28 +656,3 @@ class LeannChat:
             except (KeyboardInterrupt, EOFError):
                 print("\nGoodbye!")
                 break
-
-    def cleanup(self):
-        """Explicitly cleanup embedding server resources.
-
-        This method should be called after you're done using the chat interface,
-        especially in test environments or batch processing scenarios.
-        """
-        if hasattr(self.searcher, "cleanup"):
-            self.searcher.cleanup()
-
-    # Enable automatic cleanup patterns
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc, tb):
-        try:
-            self.cleanup()
-        except Exception:
-            pass
-
-    def __del__(self):
-        try:
-            self.cleanup()
-        except Exception:
-            pass

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

@@ -8,7 +8,7 @@ import difflib
 import logging
 import os
 from abc import ABC, abstractmethod
-from typing import Any, Optional
+from typing import Any
 
 import torch
 
@@ -17,12 +17,12 @@ logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
 
-def check_ollama_models(host: str) -> list[str]:
+def check_ollama_models() -> list[str]:
     """Check available Ollama models and return a list"""
     try:
         import requests
 
-        response = requests.get(f"{host}/api/tags", timeout=5)
+        response = requests.get("http://localhost:11434/api/tags", timeout=5)
         if response.status_code == 200:
             data = response.json()
             return [model["name"] for model in data.get("models", [])]
@@ -309,12 +309,10 @@ def search_hf_models(query: str, limit: int = 10) -> list[str]:
     return search_hf_models_fuzzy(query, limit)
 
 
-def validate_model_and_suggest(
+def validate_model_and_suggest(model_name: str, llm_type: str) -> str | None:
-    model_name: str, llm_type: str, host: str = "http://localhost:11434"
-) -> Optional[str]:
     """Validate model name and provide suggestions if invalid"""
     if llm_type == "ollama":
-        available_models = check_ollama_models(host)
+        available_models = check_ollama_models()
         if available_models and model_name not in available_models:
             error_msg = f"Model '{model_name}' not found in your local Ollama installation."
 
@@ -422,6 +420,7 @@ class LLMInterface(ABC):
                 top_k=10,
                 complexity=64,
                 beam_width=8,
+                USE_DEFERRED_FETCH=True,
                 skip_search_reorder=True,
                 recompute_beighbor_embeddings=True,
                 dedup_node_dis=True,
@@ -433,6 +432,7 @@ class LLMInterface(ABC):
         Supported kwargs:
             - complexity (int): Search complexity parameter (default: 32)
             - beam_width (int): Beam width for search (default: 4)
+            - USE_DEFERRED_FETCH (bool): Enable deferred fetch mode (default: False)
             - skip_search_reorder (bool): Skip search reorder step (default: False)
             - recompute_beighbor_embeddings (bool): Enable ZMQ embedding server for neighbor recomputation (default: False)
             - dedup_node_dis (bool): Deduplicate nodes by distance (default: False)
@@ -469,7 +469,7 @@ class OllamaChat(LLMInterface):
                 requests.get(host)
 
             # Pre-check model availability with helpful suggestions
-            model_error = validate_model_and_suggest(model, "ollama", host)
+            model_error = validate_model_and_suggest(model, "ollama")
             if model_error:
                 raise ValueError(model_error)
 
@@ -489,35 +489,11 @@ class OllamaChat(LLMInterface):
         import requests
 
         full_url = f"{self.host}/api/generate"
-
-        # Handle thinking budget for reasoning models
-        options = kwargs.copy()
-        thinking_budget = kwargs.get("thinking_budget")
-        if thinking_budget:
-            # Remove thinking_budget from options as it's not a standard Ollama option
-            options.pop("thinking_budget", None)
-            # Only apply reasoning parameters to models that support it
-            reasoning_supported_models = [
-                "gpt-oss:20b",
-                "gpt-oss:120b",
-                "deepseek-r1",
-                "deepseek-coder",
-            ]
-
-            if thinking_budget in ["low", "medium", "high"]:
-                if any(model in self.model.lower() for model in reasoning_supported_models):
-                    options["reasoning"] = {"effort": thinking_budget, "exclude": False}
-                    logger.info(f"Applied reasoning effort={thinking_budget} to model {self.model}")
-                else:
-                    logger.warning(
-                        f"Thinking budget '{thinking_budget}' requested but model '{self.model}' may not support reasoning parameters. Proceeding without reasoning."
-                    )
-
         payload = {
             "model": self.model,
             "prompt": prompt,
             "stream": False,  # Keep it simple for now
-            "options": options,
+            "options": kwargs,
         }
         logger.debug(f"Sending request to Ollama: {payload}")
         try:
@@ -683,7 +659,7 @@ class HFChat(LLMInterface):
 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")
 
@@ -708,38 +684,11 @@ class OpenAIChat(LLMInterface):
         params = {
             "model": self.model,
             "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"]},
         }
 
-        # Handle max_tokens vs max_completion_tokens based on model
-        max_tokens = kwargs.get("max_tokens", 1000)
-        if "o3" in self.model or "o4" in self.model or "o1" in self.model:
-            # o-series models use max_completion_tokens
-            params["max_completion_tokens"] = max_tokens
-            params["temperature"] = 1.0
-        else:
-            # Other models use max_tokens
-            params["max_tokens"] = max_tokens
-
-        # Handle thinking budget for reasoning models
-        thinking_budget = kwargs.get("thinking_budget")
-        if thinking_budget and thinking_budget in ["low", "medium", "high"]:
-            # Check if this is an o-series model (partial match for model names)
-            o_series_models = ["o3", "o3-mini", "o4-mini", "o1", "o3-pro", "o3-deep-research"]
-            if any(model in self.model for model in o_series_models):
-                # Use the correct OpenAI reasoning parameter format
-                params["reasoning_effort"] = thinking_budget
-                logger.info(f"Applied reasoning_effort={thinking_budget} to model {self.model}")
-            else:
-                logger.warning(
-                    f"Thinking budget '{thinking_budget}' requested but model '{self.model}' may not support reasoning parameters. Proceeding without reasoning."
-                )
-
-        # Add other kwargs (excluding thinking_budget as it's handled above)
-        for k, v in kwargs.items():
-            if k not in ["max_tokens", "temperature", "thinking_budget"]:
-                params[k] = v
-
         logger.info(f"Sending request to OpenAI with model {self.model}")
 
         try:
@@ -759,7 +708,7 @@ 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

@@ -1,11 +1,9 @@
 import argparse
 import asyncio
 from pathlib import Path
-from typing import Union
 
 from llama_index.core import SimpleDirectoryReader
 from llama_index.core.node_parser import SentenceSplitter
-from tqdm import tqdm
 
 from .api import LeannBuilder, LeannChat, LeannSearcher
 
@@ -43,23 +41,13 @@ def extract_pdf_text_with_pdfplumber(file_path: str) -> str:
 
 class LeannCLI:
     def __init__(self):
-        # Always use project-local .leann directory (like .git)
+        self.indexes_dir = Path.home() / ".leann" / "indexes"
-        self.indexes_dir = Path.cwd() / ".leann" / "indexes"
         self.indexes_dir.mkdir(parents=True, exist_ok=True)
 
-        # Default parser for documents
         self.node_parser = SentenceSplitter(
             chunk_size=256, chunk_overlap=128, separator=" ", paragraph_separator="\n\n"
         )
 
-        # Code-optimized parser
-        self.code_parser = SentenceSplitter(
-            chunk_size=512,  # Larger chunks for code context
-            chunk_overlap=50,  # Less overlap to preserve function boundaries
-            separator="\n",  # Split by lines for code
-            paragraph_separator="\n\n",  # Preserve logical code blocks
-        )
-
     def get_index_path(self, index_name: str) -> str:
         index_dir = self.indexes_dir / index_name
         return str(index_dir / "documents.leann")
@@ -72,18 +60,14 @@ class LeannCLI:
     def create_parser(self) -> argparse.ArgumentParser:
         parser = argparse.ArgumentParser(
             prog="leann",
-            description="The smallest vector index in the world. RAG Everything with LEANN!",
+            description="LEANN - Local Enhanced AI Navigation",
             formatter_class=argparse.RawDescriptionHelpFormatter,
             epilog="""
 Examples:
-  leann build my-docs --docs ./documents                                  # Build index from directory
+  leann build my-docs --docs ./documents    # Build index named my-docs
-  leann build my-code --docs ./src ./tests ./config                      # Build index from multiple directories
+  leann search my-docs "query"             # Search in my-docs index
-  leann build my-files --docs ./file1.py ./file2.txt ./docs/             # Build index from files and directories
+  leann ask my-docs "question"             # Ask my-docs index
-  leann build my-mixed --docs ./readme.md ./src/ ./config.json           # Build index from mixed files/dirs
+  leann list                              # List all stored indexes
-  leann build my-ppts --docs ./ --file-types .pptx,.pdf                  # Index only PowerPoint and PDF files
-  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
             """,
         )
 
@@ -91,88 +75,32 @@ Examples:
 
         # 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(
-            "index_name", nargs="?", help="Index name (default: current directory name)"
+            "--backend", type=str, default="hnsw", choices=["hnsw", "diskann"]
-        )
-        build_parser.add_argument(
-            "--docs",
-            type=str,
-            nargs="+",
-            default=["."],
-            help="Documents directories and/or files (default: current directory)",
-        )
-        build_parser.add_argument(
-            "--backend",
-            type=str,
-            default="hnsw",
-            choices=["hnsw", "diskann"],
-            help="Backend to use (default: hnsw)",
-        )
-        build_parser.add_argument(
-            "--embedding-model",
-            type=str,
-            default="facebook/contriever",
-            help="Embedding model (default: facebook/contriever)",
-        )
-        build_parser.add_argument(
-            "--embedding-mode",
-            type=str,
-            default="sentence-transformers",
-            choices=["sentence-transformers", "openai", "mlx", "ollama"],
-            help="Embedding backend mode (default: sentence-transformers)",
-        )
-        build_parser.add_argument(
-            "--force", "-f", action="store_true", help="Force rebuild existing index"
-        )
-        build_parser.add_argument(
-            "--graph-degree", type=int, default=32, help="Graph degree (default: 32)"
-        )
-        build_parser.add_argument(
-            "--complexity", type=int, default=64, help="Build complexity (default: 64)"
         )
+        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)
+        build_parser.add_argument("--complexity", type=int, default=64)
         build_parser.add_argument("--num-threads", type=int, default=1)
-        build_parser.add_argument(
+        build_parser.add_argument("--compact", action="store_true", default=True)
-            "--compact",
+        build_parser.add_argument("--recompute", action="store_true", default=True)
-            action=argparse.BooleanOptionalAction,
-            default=True,
-            help="Use compact storage (default: true). Must be `no-compact` for `no-recompute` build.",
-        )
-        build_parser.add_argument(
-            "--recompute",
-            action=argparse.BooleanOptionalAction,
-            default=True,
-            help="Enable recomputation (default: true)",
-        )
-        build_parser.add_argument(
-            "--file-types",
-            type=str,
-            help="Comma-separated list of file extensions to include (e.g., '.txt,.pdf,.pptx'). If not specified, uses default supported types.",
-        )
 
         # Search command
         search_parser = subparsers.add_parser("search", help="Search documents")
         search_parser.add_argument("index_name", help="Index name")
         search_parser.add_argument("query", help="Search query")
-        search_parser.add_argument(
+        search_parser.add_argument("--top-k", type=int, default=5)
-            "--top-k", type=int, default=5, help="Number of results (default: 5)"
+        search_parser.add_argument("--complexity", type=int, default=64)
-        )
-        search_parser.add_argument(
-            "--complexity", type=int, default=64, help="Search complexity (default: 64)"
-        )
         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(
+        search_parser.add_argument("--recompute-embeddings", action="store_true")
-            "--recompute",
-            dest="recompute_embeddings",
-            action=argparse.BooleanOptionalAction,
-            default=True,
-            help="Enable/disable embedding recomputation (default: enabled). Should not do a `no-recompute` search in a `recompute` build.",
-        )
         search_parser.add_argument(
             "--pruning-strategy",
             choices=["global", "local", "proportional"],
             default="global",
-            help="Pruning strategy (default: global)",
         )
 
         # Ask command
@@ -183,513 +111,102 @@ Examples:
             type=str,
             default="ollama",
             choices=["simulated", "ollama", "hf", "openai"],
-            help="LLM provider (default: ollama)",
-        )
-        ask_parser.add_argument(
-            "--model", type=str, default="qwen3:8b", help="Model name (default: qwen3:8b)"
         )
+        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(
+        ask_parser.add_argument("--interactive", "-i", action="store_true")
-            "--interactive", "-i", action="store_true", help="Interactive chat mode"
+        ask_parser.add_argument("--top-k", type=int, default=20)
-        )
-        ask_parser.add_argument(
-            "--top-k", type=int, default=20, help="Retrieval count (default: 20)"
-        )
         ask_parser.add_argument("--complexity", type=int, default=32)
         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(
+        ask_parser.add_argument("--recompute-embeddings", action="store_true")
-            "--recompute",
-            dest="recompute_embeddings",
-            action=argparse.BooleanOptionalAction,
-            default=True,
-            help="Enable/disable embedding recomputation during ask (default: enabled)",
-        )
         ask_parser.add_argument(
             "--pruning-strategy",
             choices=["global", "local", "proportional"],
             default="global",
         )
-        ask_parser.add_argument(
-            "--thinking-budget",
-            type=str,
-            choices=["low", "medium", "high"],
-            default=None,
-            help="Thinking budget for reasoning models (low/medium/high). Supported by GPT-Oss:20b and other reasoning models.",
-        )
 
         # List command
         subparsers.add_parser("list", help="List all indexes")
 
         return parser
 
-    def register_project_dir(self):
-        """Register current project directory in global registry"""
-        global_registry = Path.home() / ".leann" / "projects.json"
-        global_registry.parent.mkdir(exist_ok=True)
-
-        current_dir = str(Path.cwd())
-
-        # Load existing registry
-        projects = []
-        if global_registry.exists():
-            try:
-                import json
-
-                with open(global_registry) as f:
-                    projects = json.load(f)
-            except Exception:
-                projects = []
-
-        # Add current directory if not already present
-        if current_dir not in projects:
-            projects.append(current_dir)
-
-        # Save registry
-        import json
-
-        with open(global_registry, "w") as f:
-            json.dump(projects, f, indent=2)
-
-    def _build_gitignore_parser(self, docs_dir: str):
-        """Build gitignore parser using gitignore-parser library."""
-        from gitignore_parser import parse_gitignore
-
-        # Try to parse the root .gitignore
-        gitignore_path = Path(docs_dir) / ".gitignore"
-
-        if gitignore_path.exists():
-            try:
-                # gitignore-parser automatically handles all subdirectory .gitignore files!
-                matches = parse_gitignore(str(gitignore_path))
-                print(f"📋 Loaded .gitignore from {docs_dir} (includes all subdirectories)")
-                return matches
-            except Exception as e:
-                print(f"Warning: Could not parse .gitignore: {e}")
-        else:
-            print("📋 No .gitignore found")
-
-        # Fallback: basic pattern matching for essential files
-        essential_patterns = {".git", ".DS_Store", "__pycache__", "node_modules", ".venv", "venv"}
-
-        def basic_matches(file_path):
-            path_parts = Path(file_path).parts
-            return any(part in essential_patterns for part in path_parts)
-
-        return basic_matches
-
-    def _should_exclude_file(self, relative_path: Path, gitignore_matches) -> bool:
-        """Check if a file should be excluded using gitignore parser."""
-        return gitignore_matches(str(relative_path))
-
-    def _is_git_submodule(self, path: Path) -> bool:
-        """Check if a path is a git submodule."""
-        try:
-            # Find the git repo root
-            current_dir = Path.cwd()
-            while current_dir != current_dir.parent:
-                if (current_dir / ".git").exists():
-                    gitmodules_path = current_dir / ".gitmodules"
-                    if gitmodules_path.exists():
-                        # Read .gitmodules to check if this path is a submodule
-                        gitmodules_content = gitmodules_path.read_text()
-                        # Convert path to relative to git root
-                        try:
-                            relative_path = path.resolve().relative_to(current_dir)
-                            # Check if this path appears in .gitmodules
-                            return f"path = {relative_path}" in gitmodules_content
-                        except ValueError:
-                            # Path is not under git root
-                            return False
-                    break
-                current_dir = current_dir.parent
-            return False
-        except Exception:
-            # If anything goes wrong, assume it's not a submodule
-            return False
-
     def list_indexes(self):
         print("Stored LEANN indexes:")
 
-        # Get all project directories with .leann
+        if not self.indexes_dir.exists():
-        global_registry = Path.home() / ".leann" / "projects.json"
+            print("No indexes found. Use 'leann build <name> --docs <dir>' to create one.")
-        all_projects = []
-
-        if global_registry.exists():
-            try:
-                import json
-
-                with open(global_registry) as f:
-                    all_projects = json.load(f)
-            except Exception:
-                pass
-
-        # Filter to only existing directories with .leann
-        valid_projects = []
-        for project_dir in all_projects:
-            project_path = Path(project_dir)
-            if project_path.exists() and (project_path / ".leann" / "indexes").exists():
-                valid_projects.append(project_path)
-
-        # Add current project if it has .leann but not in registry
-        current_path = Path.cwd()
-        if (current_path / ".leann" / "indexes").exists() and current_path not in valid_projects:
-            valid_projects.append(current_path)
-
-        if not valid_projects:
-            print(
-                "No indexes found. Use 'leann build <name> --docs <dir> [<dir2> ...]' to create one."
-            )
             return
 
-        total_indexes = 0
+        index_dirs = [d for d in self.indexes_dir.iterdir() if d.is_dir()]
-        current_dir = Path.cwd()
 
-        for project_path in valid_projects:
+        if not index_dirs:
-            indexes_dir = project_path / ".leann" / "indexes"
+            print("No indexes found. Use 'leann build <name> --docs <dir>' to create one.")
-            if not indexes_dir.exists():
+            return
-                continue
 
-            index_dirs = [d for d in indexes_dir.iterdir() if d.is_dir()]
+        print(f"Found {len(index_dirs)} indexes:")
-            if not index_dirs:
+        for i, index_dir in enumerate(index_dirs, 1):
-                continue
+            index_name = index_dir.name
+            status = "✓" if self.index_exists(index_name) else "✗"
 
-            # Show project header
+            print(f"  {i}. {index_name} [{status}]")
-            if project_path == current_dir:
+            if self.index_exists(index_name):
-                print(f"\n📁 Current project ({project_path}):")
+                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("\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}...")
+
+        # 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:
-                print(f"\n📂 {project_path}:")
+                # Fallback to default reader
-
+                print(f"Using default reader for {file_path}")
-            for index_dir in index_dirs:
+                default_docs = SimpleDirectoryReader(
-                total_indexes += 1
+                    str(file_path.parent),
-                index_name = index_dir.name
-                meta_file = index_dir / "documents.leann.meta.json"
-                status = "✓" if meta_file.exists() else "✗"
-
-                print(f"  {total_indexes}. {index_name} [{status}]")
-                if status == "✓":
-                    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 total_indexes > 0:
-            print(f"\nTotal: {total_indexes} indexes across {len(valid_projects)} projects")
-            print("\nUsage (current project only):")
-
-            # Show example from current project
-            current_indexes_dir = current_dir / ".leann" / "indexes"
-            if current_indexes_dir.exists():
-                current_index_dirs = [d for d in current_indexes_dir.iterdir() if d.is_dir()]
-                if current_index_dirs:
-                    example_name = current_index_dirs[0].name
-                    print(f'  leann search {example_name} "your query"')
-                    print(f"  leann ask {example_name} --interactive")
-
-    def load_documents(
-        self, docs_paths: Union[str, list], custom_file_types: Union[str, None] = None
-    ):
-        # Handle both single path (string) and multiple paths (list) for backward compatibility
-        if isinstance(docs_paths, str):
-            docs_paths = [docs_paths]
-
-        # Separate files and directories
-        files = []
-        directories = []
-        for path in docs_paths:
-            path_obj = Path(path)
-            if path_obj.is_file():
-                files.append(str(path_obj))
-            elif path_obj.is_dir():
-                # Check if this is a git submodule - if so, skip it
-                if self._is_git_submodule(path_obj):
-                    print(f"⚠️  Skipping git submodule: {path}")
-                    continue
-                directories.append(str(path_obj))
-            else:
-                print(f"⚠️  Warning: Path '{path}' does not exist, skipping...")
-                continue
-
-        # Print summary of what we're processing
-        total_items = len(files) + len(directories)
-        items_desc = []
-        if files:
-            items_desc.append(f"{len(files)} file{'s' if len(files) > 1 else ''}")
-        if directories:
-            items_desc.append(
-                f"{len(directories)} director{'ies' if len(directories) > 1 else 'y'}"
-            )
-
-        print(f"Loading documents from {' and '.join(items_desc)} ({total_items} total):")
-        if files:
-            print(f"  📄 Files: {', '.join([Path(f).name for f in files])}")
-        if directories:
-            print(f"  📁 Directories: {', '.join(directories)}")
-
-        if custom_file_types:
-            print(f"Using custom file types: {custom_file_types}")
-
-        all_documents = []
-
-        # First, process individual files if any
-        if files:
-            print(f"\n🔄 Processing {len(files)} individual file{'s' if len(files) > 1 else ''}...")
-
-            # Load individual files using SimpleDirectoryReader with input_files
-            # Note: We skip gitignore filtering for explicitly specified files
-            try:
-                # Group files by their parent directory for efficient loading
-                from collections import defaultdict
-
-                files_by_dir = defaultdict(list)
-                for file_path in files:
-                    parent_dir = str(Path(file_path).parent)
-                    files_by_dir[parent_dir].append(file_path)
-
-                # Load files from each parent directory
-                for parent_dir, file_list in files_by_dir.items():
-                    print(
-                        f"  Loading {len(file_list)} file{'s' if len(file_list) > 1 else ''} from {parent_dir}"
-                    )
-                    try:
-                        file_docs = SimpleDirectoryReader(
-                            parent_dir,
-                            input_files=file_list,
-                            filename_as_id=True,
-                        ).load_data()
-                        all_documents.extend(file_docs)
-                        print(
-                            f"    ✅ Loaded {len(file_docs)} document{'s' if len(file_docs) > 1 else ''}"
-                        )
-                    except Exception as e:
-                        print(f"    ❌ Warning: Could not load files from {parent_dir}: {e}")
-
-            except Exception as e:
-                print(f"❌ Error processing individual files: {e}")
-
-        # Define file extensions to process
-        if custom_file_types:
-            # Parse custom file types from comma-separated string
-            code_extensions = [ext.strip() for ext in custom_file_types.split(",") if ext.strip()]
-            # Ensure extensions start with a dot
-            code_extensions = [ext if ext.startswith(".") else f".{ext}" for ext in code_extensions]
-        else:
-            # Use default supported file types
-            code_extensions = [
-                # Original document types
-                ".txt",
-                ".md",
-                ".docx",
-                ".pptx",
-                # Code files for Claude Code integration
-                ".py",
-                ".js",
-                ".ts",
-                ".jsx",
-                ".tsx",
-                ".java",
-                ".cpp",
-                ".c",
-                ".h",
-                ".hpp",
-                ".cs",
-                ".go",
-                ".rs",
-                ".rb",
-                ".php",
-                ".swift",
-                ".kt",
-                ".scala",
-                ".r",
-                ".sql",
-                ".sh",
-                ".bash",
-                ".zsh",
-                ".fish",
-                ".ps1",
-                ".bat",
-                # Config and markup files
-                ".json",
-                ".yaml",
-                ".yml",
-                ".xml",
-                ".toml",
-                ".ini",
-                ".cfg",
-                ".conf",
-                ".html",
-                ".css",
-                ".scss",
-                ".less",
-                ".vue",
-                ".svelte",
-                # Data science
-                ".ipynb",
-                ".R",
-                ".py",
-                ".jl",
-            ]
-
-        # Process each directory
-        if directories:
-            print(
-                f"\n🔄 Processing {len(directories)} director{'ies' if len(directories) > 1 else 'y'}..."
-            )
-
-        for docs_dir in directories:
-            print(f"Processing directory: {docs_dir}")
-            # Build gitignore parser for each directory
-            gitignore_matches = self._build_gitignore_parser(docs_dir)
-
-            # Try to use better PDF parsers first, but only if PDFs are requested
-            documents = []
-            docs_path = Path(docs_dir)
-
-            # Check if we should process PDFs
-            should_process_pdfs = custom_file_types is None or ".pdf" in custom_file_types
-
-            if should_process_pdfs:
-                for file_path in docs_path.rglob("*.pdf"):
-                    # Check if file matches any exclude pattern
-                    try:
-                        relative_path = file_path.relative_to(docs_path)
-                        if self._should_exclude_file(relative_path, gitignore_matches):
-                            continue
-                    except ValueError:
-                        # Skip files that can't be made relative to docs_path
-                        print(f"⚠️  Skipping file outside directory scope: {file_path}")
-                        continue
-
-                    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}")
-                        try:
-                            default_docs = SimpleDirectoryReader(
-                                str(file_path.parent),
-                                filename_as_id=True,
-                                required_exts=[file_path.suffix],
-                            ).load_data()
-                            documents.extend(default_docs)
-                        except Exception as e:
-                            print(f"Warning: Could not process {file_path}: {e}")
-
-            # Load other file types with default reader
-            try:
-                # Create a custom file filter function using our PathSpec
-                def file_filter(
-                    file_path: str, docs_dir=docs_dir, gitignore_matches=gitignore_matches
-                ) -> bool:
-                    """Return True if file should be included (not excluded)"""
-                    try:
-                        docs_path_obj = Path(docs_dir)
-                        file_path_obj = Path(file_path)
-                        relative_path = file_path_obj.relative_to(docs_path_obj)
-                        return not self._should_exclude_file(relative_path, gitignore_matches)
-                    except (ValueError, OSError):
-                        return True  # Include files that can't be processed
-
-                other_docs = SimpleDirectoryReader(
-                    docs_dir,
-                    recursive=True,
-                    encoding="utf-8",
-                    required_exts=code_extensions,
-                    file_extractor={},  # Use default extractors
                     filename_as_id=True,
-                ).load_data(show_progress=True)
+                    required_exts=[file_path.suffix],
+                ).load_data()
+                documents.extend(default_docs)
 
-                # Filter documents after loading based on gitignore rules
+        # Load other file types with default reader
-                filtered_docs = []
+        other_docs = SimpleDirectoryReader(
-                for doc in other_docs:
+            docs_dir,
-                    file_path = doc.metadata.get("file_path", "")
+            recursive=True,
-                    if file_filter(file_path):
+            encoding="utf-8",
-                        filtered_docs.append(doc)
+            required_exts=[".txt", ".md", ".docx"],
-
+        ).load_data(show_progress=True)
-                documents.extend(filtered_docs)
+        documents.extend(other_docs)
-            except ValueError as e:
-                if "No files found" in str(e):
-                    print(f"No additional files found for other supported types in {docs_dir}.")
-                else:
-                    raise e
-
-            all_documents.extend(documents)
-            print(f"Loaded {len(documents)} documents from {docs_dir}")
-
-        documents = all_documents
 
         all_texts = []
-
+        for doc in documents:
-        # Define code file extensions for intelligent chunking
+            nodes = self.node_parser.get_nodes_from_documents([doc])
-        code_file_exts = {
-            ".py",
-            ".js",
-            ".ts",
-            ".jsx",
-            ".tsx",
-            ".java",
-            ".cpp",
-            ".c",
-            ".h",
-            ".hpp",
-            ".cs",
-            ".go",
-            ".rs",
-            ".rb",
-            ".php",
-            ".swift",
-            ".kt",
-            ".scala",
-            ".r",
-            ".sql",
-            ".sh",
-            ".bash",
-            ".zsh",
-            ".fish",
-            ".ps1",
-            ".bat",
-            ".json",
-            ".yaml",
-            ".yml",
-            ".xml",
-            ".toml",
-            ".ini",
-            ".cfg",
-            ".conf",
-            ".html",
-            ".css",
-            ".scss",
-            ".less",
-            ".vue",
-            ".svelte",
-            ".ipynb",
-            ".R",
-            ".jl",
-        }
-
-        print("start chunking documents")
-        # Add progress bar for document chunking
-        for doc in tqdm(documents, desc="Chunking documents", unit="doc"):
-            # Check if this is a code file based on source path
-            source_path = doc.metadata.get("source", "")
-            is_code_file = any(source_path.endswith(ext) for ext in code_file_exts)
-
-            # Use appropriate parser based on file type
-            parser = self.code_parser if is_code_file else self.node_parser
-            nodes = parser.get_nodes_from_documents([doc])
-
             for node in nodes:
                 all_texts.append(node.get_content())
 
@@ -697,36 +214,16 @@ Examples:
         return all_texts
 
     async def build_index(self, args):
-        docs_paths = args.docs
+        docs_dir = args.docs
-        # Use current directory name if index_name not provided
+        index_name = args.index_name
-        if args.index_name:
-            index_name = args.index_name
-        else:
-            index_name = Path.cwd().name
-            print(f"Using current directory name as index: '{index_name}'")
-
         index_dir = self.indexes_dir / index_name
         index_path = self.get_index_path(index_name)
 
-        # Display all paths being indexed with file/directory distinction
-        files = [p for p in docs_paths if Path(p).is_file()]
-        directories = [p for p in docs_paths if Path(p).is_dir()]
-
-        print(f"📂 Indexing {len(docs_paths)} path{'s' if len(docs_paths) > 1 else ''}:")
-        if files:
-            print(f"  📄 Files ({len(files)}):")
-            for i, file_path in enumerate(files, 1):
-                print(f"    {i}. {Path(file_path).resolve()}")
-        if directories:
-            print(f"  📁 Directories ({len(directories)}):")
-            for i, dir_path in enumerate(directories, 1):
-                print(f"    {i}. {Path(dir_path).resolve()}")
-
         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_paths, args.file_types)
+        all_texts = self.load_documents(docs_dir)
         if not all_texts:
             print("No documents found")
             return
@@ -738,7 +235,6 @@ Examples:
         builder = LeannBuilder(
             backend_name=args.backend,
             embedding_model=args.embedding_model,
-            embedding_mode=args.embedding_mode,
             graph_degree=args.graph_degree,
             complexity=args.complexity,
             is_compact=args.compact,
@@ -752,9 +248,6 @@ Examples:
         builder.build_index(index_path)
         print(f"Index built at {index_path}")
 
-        # Register this project directory in global registry
-        self.register_project_dir()
-
     async def search_documents(self, args):
         index_name = args.index_name
         query = args.query
@@ -762,7 +255,7 @@ Examples:
 
         if not self.index_exists(index_name):
             print(
-                f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir> [<dir2> ...]' to create it."
+                f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir>' to create it."
             )
             return
 
@@ -789,7 +282,7 @@ Examples:
 
         if not self.index_exists(index_name):
             print(
-                f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir> [<dir2> ...]' to create it."
+                f"Index '{index_name}' not found. Use 'leann build {index_name} --docs <dir>' to create it."
             )
             return
 
@@ -815,11 +308,6 @@ Examples:
                 if not user_input:
                     continue
 
-                # Prepare LLM kwargs with thinking budget if specified
-                llm_kwargs = {}
-                if args.thinking_budget:
-                    llm_kwargs["thinking_budget"] = args.thinking_budget
-
                 response = chat.ask(
                     user_input,
                     top_k=args.top_k,
@@ -828,17 +316,11 @@ Examples:
                     prune_ratio=args.prune_ratio,
                     recompute_embeddings=args.recompute_embeddings,
                     pruning_strategy=args.pruning_strategy,
-                    llm_kwargs=llm_kwargs,
                 )
                 print(f"LEANN: {response}")
         else:
             query = input("Enter your question: ").strip()
             if query:
-                # Prepare LLM kwargs with thinking budget if specified
-                llm_kwargs = {}
-                if args.thinking_budget:
-                    llm_kwargs["thinking_budget"] = args.thinking_budget
-
                 response = chat.ask(
                     query,
                     top_k=args.top_k,
@@ -847,7 +329,6 @@ Examples:
                     prune_ratio=args.prune_ratio,
                     recompute_embeddings=args.recompute_embeddings,
                     pruning_strategy=args.pruning_strategy,
-                    llm_kwargs=llm_kwargs,
                 )
                 print(f"LEANN: {response}")
 

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

@@ -35,7 +35,7 @@ def compute_embeddings(
     Args:
         texts: List of texts to compute embeddings for
         model_name: Model name
-        mode: Computation mode ('sentence-transformers', 'openai', 'mlx', 'ollama')
+        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
@@ -55,8 +55,6 @@ def compute_embeddings(
         return compute_embeddings_openai(texts, model_name)
     elif mode == "mlx":
         return compute_embeddings_mlx(texts, model_name)
-    elif mode == "ollama":
-        return compute_embeddings_ollama(texts, model_name, is_build=is_build)
     else:
         raise ValueError(f"Unsupported embedding mode: {mode}")
 
@@ -367,286 +365,3 @@ def compute_embeddings_mlx(chunks: list[str], model_name: str, batch_size: int =
 
     # Stack numpy arrays
     return np.stack(all_embeddings)
-
-
-def compute_embeddings_ollama(
-    texts: list[str], model_name: str, is_build: bool = False, host: str = "http://localhost:11434"
-) -> np.ndarray:
-    """
-    Compute embeddings using Ollama API with simplified batch processing.
-
-    Uses batch size of 32 for MPS/CPU and 128 for CUDA to optimize performance.
-
-    Args:
-        texts: List of texts to compute embeddings for
-        model_name: Ollama model name (e.g., "nomic-embed-text", "mxbai-embed-large")
-        is_build: Whether this is a build operation (shows progress bar)
-        host: Ollama host URL (default: http://localhost:11434)
-
-    Returns:
-        Normalized embeddings array, shape: (len(texts), embedding_dim)
-    """
-    try:
-        import requests
-    except ImportError:
-        raise ImportError(
-            "The 'requests' library is required for Ollama embeddings. Install with: uv pip install requests"
-        )
-
-    if not texts:
-        raise ValueError("Cannot compute embeddings for empty text list")
-
-    logger.info(
-        f"Computing embeddings for {len(texts)} texts using Ollama API, model: '{model_name}'"
-    )
-
-    # Check if Ollama is running
-    try:
-        response = requests.get(f"{host}/api/version", timeout=5)
-        response.raise_for_status()
-    except requests.exceptions.ConnectionError:
-        error_msg = (
-            f"❌ Could not connect to Ollama at {host}.\n\n"
-            "Please ensure Ollama is running:\n"
-            "  • macOS/Linux: ollama serve\n"
-            "  • Windows: Make sure Ollama is running in the system tray\n\n"
-            "Installation: https://ollama.com/download"
-        )
-        raise RuntimeError(error_msg)
-    except Exception as e:
-        raise RuntimeError(f"Unexpected error connecting to Ollama: {e}")
-
-    # Check if model exists and provide helpful suggestions
-    try:
-        response = requests.get(f"{host}/api/tags", timeout=5)
-        response.raise_for_status()
-        models = response.json()
-        model_names = [model["name"] for model in models.get("models", [])]
-
-        # Filter for embedding models (models that support embeddings)
-        embedding_models = []
-        suggested_embedding_models = [
-            "nomic-embed-text",
-            "mxbai-embed-large",
-            "bge-m3",
-            "all-minilm",
-            "snowflake-arctic-embed",
-        ]
-
-        for model in model_names:
-            # Check if it's an embedding model (by name patterns or known models)
-            base_name = model.split(":")[0]
-            if any(emb in base_name for emb in ["embed", "bge", "minilm", "e5"]):
-                embedding_models.append(model)
-
-        # Check if model exists (handle versioned names) and resolve to full name
-        resolved_model_name = None
-        for name in model_names:
-            # Exact match
-            if model_name == name:
-                resolved_model_name = name
-                break
-            # Match without version tag (use the versioned name)
-            elif model_name == name.split(":")[0]:
-                resolved_model_name = name
-                break
-
-        if not resolved_model_name:
-            error_msg = f"❌ Model '{model_name}' not found in local Ollama.\n\n"
-
-            # Suggest pulling the model
-            error_msg += "📦 To install this embedding model:\n"
-            error_msg += f"   ollama pull {model_name}\n\n"
-
-            # Show available embedding models
-            if embedding_models:
-                error_msg += "✅ Available embedding models:\n"
-                for model in embedding_models[:5]:
-                    error_msg += f"   • {model}\n"
-                if len(embedding_models) > 5:
-                    error_msg += f"   ... and {len(embedding_models) - 5} more\n"
-            else:
-                error_msg += "💡 Popular embedding models to install:\n"
-                for model in suggested_embedding_models[:3]:
-                    error_msg += f"   • ollama pull {model}\n"
-
-            error_msg += "\n📚 Browse more: https://ollama.com/library"
-            raise ValueError(error_msg)
-
-        # Use the resolved model name for all subsequent operations
-        if resolved_model_name != model_name:
-            logger.info(f"Resolved model name '{model_name}' to '{resolved_model_name}'")
-        model_name = resolved_model_name
-
-        # Verify the model supports embeddings by testing it
-        try:
-            test_response = requests.post(
-                f"{host}/api/embeddings", json={"model": model_name, "prompt": "test"}, timeout=10
-            )
-            if test_response.status_code != 200:
-                error_msg = (
-                    f"⚠️ Model '{model_name}' exists but may not support embeddings.\n\n"
-                    f"Please use an embedding model like:\n"
-                )
-                for model in suggested_embedding_models[:3]:
-                    error_msg += f"   • {model}\n"
-                raise ValueError(error_msg)
-        except requests.exceptions.RequestException:
-            # If test fails, continue anyway - model might still work
-            pass
-
-    except requests.exceptions.RequestException as e:
-        logger.warning(f"Could not verify model existence: {e}")
-
-    # Determine batch size based on device availability
-    # Check for CUDA/MPS availability using torch if available
-    batch_size = 32  # Default for MPS/CPU
-    try:
-        import torch
-
-        if torch.cuda.is_available():
-            batch_size = 128  # CUDA gets larger batch size
-        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
-            batch_size = 32  # MPS gets smaller batch size
-    except ImportError:
-        # If torch is not available, use conservative batch size
-        batch_size = 32
-
-    logger.info(f"Using batch size: {batch_size}")
-
-    def get_batch_embeddings(batch_texts):
-        """Get embeddings for a batch of texts."""
-        all_embeddings = []
-        failed_indices = []
-
-        for i, text in enumerate(batch_texts):
-            max_retries = 3
-            retry_count = 0
-
-            # Truncate very long texts to avoid API issues
-            truncated_text = text[:8000] if len(text) > 8000 else text
-            while retry_count < max_retries:
-                try:
-                    response = requests.post(
-                        f"{host}/api/embeddings",
-                        json={"model": model_name, "prompt": truncated_text},
-                        timeout=30,
-                    )
-                    response.raise_for_status()
-
-                    result = response.json()
-                    embedding = result.get("embedding")
-
-                    if embedding is None:
-                        raise ValueError(f"No embedding returned for text {i}")
-
-                    if not isinstance(embedding, list) or len(embedding) == 0:
-                        raise ValueError(f"Invalid embedding format for text {i}")
-
-                    all_embeddings.append(embedding)
-                    break
-
-                except requests.exceptions.Timeout:
-                    retry_count += 1
-                    if retry_count >= max_retries:
-                        logger.warning(f"Timeout for text {i} after {max_retries} retries")
-                        failed_indices.append(i)
-                        all_embeddings.append(None)
-                        break
-
-                except Exception as e:
-                    retry_count += 1
-                    if retry_count >= max_retries:
-                        logger.error(f"Failed to get embedding for text {i}: {e}")
-                        failed_indices.append(i)
-                        all_embeddings.append(None)
-                        break
-        return all_embeddings, failed_indices
-
-    # Process texts in batches
-    all_embeddings = []
-    all_failed_indices = []
-
-    # Setup progress bar if needed
-    show_progress = is_build or len(texts) > 10
-    try:
-        if show_progress:
-            from tqdm import tqdm
-    except ImportError:
-        show_progress = False
-
-    # Process batches
-    num_batches = (len(texts) + batch_size - 1) // batch_size
-
-    if show_progress:
-        batch_iterator = tqdm(range(num_batches), desc="Computing Ollama embeddings")
-    else:
-        batch_iterator = range(num_batches)
-
-    for batch_idx in batch_iterator:
-        start_idx = batch_idx * batch_size
-        end_idx = min(start_idx + batch_size, len(texts))
-        batch_texts = texts[start_idx:end_idx]
-
-        batch_embeddings, batch_failed = get_batch_embeddings(batch_texts)
-
-        # Adjust failed indices to global indices
-        global_failed = [start_idx + idx for idx in batch_failed]
-        all_failed_indices.extend(global_failed)
-        all_embeddings.extend(batch_embeddings)
-
-    # Handle failed embeddings
-    if all_failed_indices:
-        if len(all_failed_indices) == len(texts):
-            raise RuntimeError("Failed to compute any embeddings")
-
-        logger.warning(
-            f"Failed to compute embeddings for {len(all_failed_indices)}/{len(texts)} texts"
-        )
-
-        # Use zero embeddings as fallback for failed ones
-        valid_embedding = next((e for e in all_embeddings if e is not None), None)
-        if valid_embedding:
-            embedding_dim = len(valid_embedding)
-            for i, embedding in enumerate(all_embeddings):
-                if embedding is None:
-                    all_embeddings[i] = [0.0] * embedding_dim
-
-    # Remove None values
-    all_embeddings = [e for e in all_embeddings if e is not None]
-
-    if not all_embeddings:
-        raise RuntimeError("No valid embeddings were computed")
-
-    # Validate embedding dimensions
-    expected_dim = len(all_embeddings[0])
-    inconsistent_dims = []
-    for i, embedding in enumerate(all_embeddings):
-        if len(embedding) != expected_dim:
-            inconsistent_dims.append((i, len(embedding)))
-
-    if inconsistent_dims:
-        error_msg = f"Ollama returned inconsistent embedding dimensions. Expected {expected_dim}, but got:\n"
-        for idx, dim in inconsistent_dims[:10]:  # Show first 10 inconsistent ones
-            error_msg += f"  - Text {idx}: {dim} dimensions\n"
-        if len(inconsistent_dims) > 10:
-            error_msg += f"  ... and {len(inconsistent_dims) - 10} more\n"
-        error_msg += f"\nThis is likely an Ollama API bug with model '{model_name}'. Please try:\n"
-        error_msg += "1. Restart Ollama service: 'ollama serve'\n"
-        error_msg += f"2. Re-pull the model: 'ollama pull {model_name}'\n"
-        error_msg += (
-            "3. Use sentence-transformers instead: --embedding-mode sentence-transformers\n"
-        )
-        error_msg += "4. Report this issue to Ollama: https://github.com/ollama/ollama/issues"
-        raise ValueError(error_msg)
-
-    # Convert to numpy array and normalize
-    embeddings = np.array(all_embeddings, dtype=np.float32)
-
-    # Normalize embeddings (L2 normalization)
-    norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
-    embeddings = embeddings / (norms + 1e-8)  # Add small epsilon to avoid division by zero
-
-    logger.info(f"Generated {len(embeddings)} embeddings, dimension: {embeddings.shape[1]}")
-
-    return embeddings

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

@@ -6,9 +6,8 @@ import subprocess
 import sys
 import time
 from pathlib import Path
-from typing import Optional
 
-# Lightweight, self-contained server manager with no cross-process inspection
+import psutil
 
 # Set up logging based on environment variable
 LOG_LEVEL = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
@@ -43,7 +42,130 @@ def _check_port(port: int) -> bool:
         return s.connect_ex(("localhost", port)) == 0
 
 
-# Note: All cross-process scanning helpers removed for simplicity
+def _check_process_matches_config(
+    port: int, expected_model: str, expected_passages_file: str
+) -> bool:
+    """
+    Check if the process using the port matches our expected model and passages file.
+    Returns True if matches, False otherwise.
+    """
+    try:
+        for proc in psutil.process_iter(["pid", "cmdline"]):
+            if not _is_process_listening_on_port(proc, port):
+                continue
+
+            cmdline = proc.info["cmdline"]
+            if not cmdline:
+                continue
+
+            return _check_cmdline_matches_config(
+                cmdline, port, expected_model, expected_passages_file
+            )
+
+        logger.debug(f"No process found listening on port {port}")
+        return False
+
+    except Exception as e:
+        logger.warning(f"Could not check process on port {port}: {e}")
+        return False
+
+
+def _is_process_listening_on_port(proc, port: int) -> bool:
+    """Check if a process is listening on the given port."""
+    try:
+        connections = proc.net_connections()
+        for conn in connections:
+            if conn.laddr.port == port and conn.status == psutil.CONN_LISTEN:
+                return True
+        return False
+    except (psutil.NoSuchProcess, psutil.AccessDenied, psutil.ZombieProcess):
+        return False
+
+
+def _check_cmdline_matches_config(
+    cmdline: list, port: int, expected_model: str, expected_passages_file: str
+) -> bool:
+    """Check if command line matches our expected configuration."""
+    cmdline_str = " ".join(cmdline)
+    logger.debug(f"Found process on port {port}: {cmdline_str}")
+
+    # Check if it's our embedding server
+    is_embedding_server = any(
+        server_type in cmdline_str
+        for server_type in [
+            "embedding_server",
+            "leann_backend_diskann.embedding_server",
+            "leann_backend_hnsw.hnsw_embedding_server",
+        ]
+    )
+
+    if not is_embedding_server:
+        logger.debug(f"Process on port {port} is not our embedding server")
+        return False
+
+    # Check model name
+    model_matches = _check_model_in_cmdline(cmdline, expected_model)
+
+    # Check passages file if provided
+    passages_matches = _check_passages_in_cmdline(cmdline, expected_passages_file)
+
+    result = model_matches and passages_matches
+    logger.debug(
+        f"model_matches: {model_matches}, passages_matches: {passages_matches}, overall: {result}"
+    )
+    return result
+
+
+def _check_model_in_cmdline(cmdline: list, expected_model: str) -> bool:
+    """Check if the command line contains the expected model."""
+    if "--model-name" not in cmdline:
+        return False
+
+    model_idx = cmdline.index("--model-name")
+    if model_idx + 1 >= len(cmdline):
+        return False
+
+    actual_model = cmdline[model_idx + 1]
+    return actual_model == expected_model
+
+
+def _check_passages_in_cmdline(cmdline: list, expected_passages_file: str) -> bool:
+    """Check if the command line contains the expected passages file."""
+    if "--passages-file" not in cmdline:
+        return False  # Expected but not found
+
+    passages_idx = cmdline.index("--passages-file")
+    if passages_idx + 1 >= len(cmdline):
+        return False
+
+    actual_passages = cmdline[passages_idx + 1]
+    expected_path = Path(expected_passages_file).resolve()
+    actual_path = Path(actual_passages).resolve()
+    return actual_path == expected_path
+
+
+def _find_compatible_port_or_next_available(
+    start_port: int, model_name: str, passages_file: str, max_attempts: int = 100
+) -> tuple[int, bool]:
+    """
+    Find a port that either has a compatible server or is available.
+    Returns (port, is_compatible) where is_compatible indicates if we found a matching server.
+    """
+    for port in range(start_port, start_port + max_attempts):
+        if not _check_port(port):
+            # Port is available
+            return port, False
+
+        # Port is in use, check if it's compatible
+        if _check_process_matches_config(port, model_name, passages_file):
+            logger.info(f"Found compatible server on port {port}")
+            return port, True
+        else:
+            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}"
+    )
 
 
 class EmbeddingServerManager:
@@ -60,18 +182,9 @@ 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_process: subprocess.Popen | None = None
-        self.server_port: Optional[int] = None
+        self.server_port: int | None = None
-        # Track last-started config for in-process reuse only
-        self._server_config: Optional[dict] = None
         self._atexit_registered = False
-        # Also register a weakref finalizer to ensure cleanup when manager is GC'ed
-        try:
-            import weakref
-
-            self._finalizer = weakref.finalize(self, self._finalize_process)
-        except Exception:
-            self._finalizer = None
 
     def start_server(
         self,
@@ -81,24 +194,26 @@ class EmbeddingServerManager:
         **kwargs,
     ) -> tuple[bool, int]:
         """Start the embedding server."""
-        # passages_file may be present in kwargs for server CLI, but we don't need it here
+        passages_file = kwargs.get("passages_file")
 
-        # If this manager already has a live server, just reuse it
+        # Check if we have a compatible server already running
-        if self.server_process and self.server_process.poll() is None and self.server_port:
+        if self._has_compatible_running_server(model_name, passages_file):
-            logger.info("Reusing in-process server")
+            logger.info("Found compatible running server!")
-            return True, self.server_port
+            return True, 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)
 
-        # Always pick a fresh available port
+        # Find a compatible port or next available
-        try:
+        actual_port, is_compatible = _find_compatible_port_or_next_available(
-            actual_port = _get_available_port(port)
+            port, model_name, passages_file
-        except RuntimeError:
+        )
-            logger.error("No available ports found")
+
-            return False, port
+        if is_compatible:
+            logger.info(f"Found compatible server on port {actual_port}")
+            return True, actual_port
 
         # Start a new server
         return self._start_new_server(actual_port, model_name, embedding_mode, **kwargs)
@@ -131,7 +246,17 @@ class EmbeddingServerManager:
             logger.error(f"Failed to start embedding server in Colab: {e}")
             return False, actual_port
 
-    # Note: No compatibility check needed; manager is per-searcher and configs are stable per instance
+    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):
+            return False
+
+        if _check_process_matches_config(self.server_port, model_name, passages_file):
+            logger.info(f"Existing server process (PID {self.server_process.pid}) is compatible")
+            return True
+
+        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
@@ -178,61 +303,22 @@ class EmbeddingServerManager:
         project_root = Path(__file__).parent.parent.parent.parent.parent
         logger.info(f"Command: {' '.join(command)}")
 
-        # In CI environment, redirect stdout to avoid buffer deadlock but keep stderr for debugging
+        # Let server output go directly to console
-        # Embedding servers use many print statements that can fill stdout buffers
+        # The server will respect LEANN_LOG_LEVEL environment variable
-        is_ci = os.environ.get("CI") == "true"
-        if is_ci:
-            stdout_target = subprocess.DEVNULL
-            stderr_target = None  # Keep stderr for error debugging in CI
-            logger.info(
-                "CI environment detected, redirecting embedding server stdout to DEVNULL, keeping stderr"
-            )
-        else:
-            stdout_target = None  # Direct to console for visible logs
-            stderr_target = None  # Direct to console for visible logs
-
-        # Start embedding server subprocess
         self.server_process = subprocess.Popen(
             command,
             cwd=project_root,
-            stdout=stdout_target,
+            stdout=None,  # Direct to console
-            stderr=stderr_target,
+            stderr=None,  # Direct to console
         )
         self.server_port = port
-        # Record config for in-process reuse
-        try:
-            self._server_config = {
-                "model_name": command[command.index("--model-name") + 1]
-                if "--model-name" in command
-                else "",
-                "passages_file": command[command.index("--passages-file") + 1]
-                if "--passages-file" in command
-                else "",
-                "embedding_mode": command[command.index("--embedding-mode") + 1]
-                if "--embedding-mode" in command
-                else "sentence-transformers",
-            }
-        except Exception:
-            self._server_config = {
-                "model_name": "",
-                "passages_file": "",
-                "embedding_mode": "sentence-transformers",
-            }
         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:
-            # Always attempt best-effort finalize at interpreter exit
+            # Use a lambda to avoid issues with bound methods
-            atexit.register(self._finalize_process)
+            atexit.register(lambda: self.stop_server() if self.server_process else None)
             self._atexit_registered = True
-        # Touch finalizer so it knows there is a live process
-        if getattr(self, "_finalizer", None) is not None and not self._finalizer.alive:
-            try:
-                import weakref
-
-                self._finalizer = weakref.finalize(self, self._finalize_process)
-            except Exception:
-                pass
 
     def _wait_for_server_ready(self, port: int) -> tuple[bool, int]:
         """Wait for the server to be ready."""
@@ -257,35 +343,24 @@ class EmbeddingServerManager:
         if not self.server_process:
             return
 
-        if self.server_process and self.server_process.poll() is not None:
+        if self.server_process.poll() is not None:
             # Process already terminated
             self.server_process = None
-            self.server_port = None
-            self._server_config = None
             return
 
         logger.info(
             f"Terminating server process (PID: {self.server_process.pid}) for backend {self.backend_module_name}..."
         )
-
+        self.server_process.terminate()
-        # Use simple termination first; if the server installed signal handlers,
-        # it will exit cleanly. Otherwise escalate to kill after a short wait.
-        try:
-            self.server_process.terminate()
-        except Exception:
-            pass
 
         try:
-            self.server_process.wait(timeout=5)  # Give more time for graceful shutdown
+            self.server_process.wait(timeout=3)
-            logger.info(f"Server process {self.server_process.pid} terminated gracefully.")
+            logger.info(f"Server process {self.server_process.pid} terminated.")
         except subprocess.TimeoutExpired:
             logger.warning(
-                f"Server process {self.server_process.pid} did not terminate within 5 seconds, force killing..."
+                f"Server process {self.server_process.pid} did not terminate gracefully within 3 seconds, killing it."
             )
-            try:
+            self.server_process.kill()
-                self.server_process.kill()
-            except Exception:
-                pass
             try:
                 self.server_process.wait(timeout=2)
                 logger.info(f"Server process {self.server_process.pid} killed successfully.")
@@ -293,33 +368,15 @@ class EmbeddingServerManager:
                 logger.error(
                     f"Failed to kill server process {self.server_process.pid} - it may be hung"
                 )
+                # Don't hang indefinitely
 
-        # Clean up process resources with timeout to avoid CI hang
+        # Clean up process resources to prevent resource tracker warnings
         try:
-            # Use shorter timeout in CI environments
+            self.server_process.wait()  # Ensure process is fully cleaned up
-            is_ci = os.environ.get("CI") == "true"
-            timeout = 3 if is_ci else 10
-            self.server_process.wait(timeout=timeout)
-            logger.info(f"Server process {self.server_process.pid} cleanup completed")
-        except subprocess.TimeoutExpired:
-            logger.warning(f"Process cleanup timeout after {timeout}s, proceeding anyway")
-        except Exception as e:
-            logger.warning(f"Error during process cleanup: {e}")
-        finally:
-            self.server_process = None
-            self.server_port = None
-            self._server_config = None
-
-    def _finalize_process(self) -> None:
-        """Best-effort cleanup used by weakref.finalize/atexit."""
-        try:
-            self.stop_server()
         except Exception:
             pass
 
-    def _adopt_existing_server(self, *args, **kwargs) -> None:
+        self.server_process = None
-        # Removed: cross-process adoption no longer supported
-        return
 
     def _launch_server_process_colab(self, command: list, port: int) -> None:
         """Launch the server process with Colab-specific settings."""
@@ -335,16 +392,10 @@ class EmbeddingServerManager:
         self.server_port = port
         logger.info(f"Colab server process started with PID: {self.server_process.pid}")
 
-        # Register atexit callback (unified)
+        # Register atexit callback
         if not self._atexit_registered:
-            atexit.register(self._finalize_process)
+            atexit.register(lambda: self.stop_server() if self.server_process else None)
             self._atexit_registered = True
-        # Record config for in-process reuse is best-effort in Colab mode
-        self._server_config = {
-            "model_name": "",
-            "passages_file": "",
-            "embedding_mode": "sentence-transformers",
-        }
 
     def _wait_for_server_ready_colab(self, port: int) -> tuple[bool, int]:
         """Wait for the server to be ready with Colab-specific timeout."""

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

@@ -1,5 +1,5 @@
 from abc import ABC, abstractmethod
-from typing import Any, Literal, Optional
+from typing import Any, Literal
 
 import numpy as np
 
@@ -34,9 +34,7 @@ class LeannBackendSearcherInterface(ABC):
         pass
 
     @abstractmethod
-    def _ensure_server_running(
+    def _ensure_server_running(self, passages_source_file: str, port: int | None, **kwargs) -> int:
-        self, passages_source_file: str, port: Optional[int], **kwargs
-    ) -> int:
         """Ensure server is running"""
         pass
 
@@ -50,7 +48,7 @@ class LeannBackendSearcherInterface(ABC):
         prune_ratio: float = 0.0,
         recompute_embeddings: bool = False,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: Optional[int] = None,
+        zmq_port: int | None = None,
         **kwargs,
     ) -> dict[str, Any]:
         """Search for nearest neighbors
@@ -76,7 +74,7 @@ class LeannBackendSearcherInterface(ABC):
         self,
         query: str,
         use_server_if_available: bool = True,
-        zmq_port: Optional[int] = None,
+        zmq_port: int | None = None,
     ) -> np.ndarray:
         """Compute embedding for a query string
 

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

@@ -1,175 +0,0 @@
-#!/usr/bin/env python3
-
-import json
-import subprocess
-import sys
-
-
-def handle_request(request):
-    if request.get("method") == "initialize":
-        return {
-            "jsonrpc": "2.0",
-            "id": request.get("id"),
-            "result": {
-                "capabilities": {"tools": {}},
-                "protocolVersion": "2024-11-05",
-                "serverInfo": {"name": "leann-mcp", "version": "1.0.0"},
-            },
-        }
-
-    elif request.get("method") == "tools/list":
-        return {
-            "jsonrpc": "2.0",
-            "id": request.get("id"),
-            "result": {
-                "tools": [
-                    {
-                        "name": "leann_search",
-                        "description": """🔍 Search code using natural language - like having a coding assistant who knows your entire codebase!
-
-🎯 **Perfect for**:
-- "How does authentication work?" → finds auth-related code
-- "Error handling patterns" → locates try-catch blocks and error logic
-- "Database connection setup" → finds DB initialization code
-- "API endpoint definitions" → locates route handlers
-- "Configuration management" → finds config files and usage
-
-💡 **Pro tip**: Use this before making any changes to understand existing patterns and conventions.""",
-                        "inputSchema": {
-                            "type": "object",
-                            "properties": {
-                                "index_name": {
-                                    "type": "string",
-                                    "description": "Name of the LEANN index to search. Use 'leann_list' first to see available indexes.",
-                                },
-                                "query": {
-                                    "type": "string",
-                                    "description": "Search query - can be natural language (e.g., 'how to handle errors') or technical terms (e.g., 'async function definition')",
-                                },
-                                "top_k": {
-                                    "type": "integer",
-                                    "default": 5,
-                                    "minimum": 1,
-                                    "maximum": 20,
-                                    "description": "Number of search results to return. Use 5-10 for focused results, 15-20 for comprehensive exploration.",
-                                },
-                                "complexity": {
-                                    "type": "integer",
-                                    "default": 32,
-                                    "minimum": 16,
-                                    "maximum": 128,
-                                    "description": "Search complexity level. Use 16-32 for fast searches (recommended), 64+ for higher precision when needed.",
-                                },
-                            },
-                            "required": ["index_name", "query"],
-                        },
-                    },
-                    {
-                        "name": "leann_status",
-                        "description": "📊 Check the health and stats of your code indexes - like a medical checkup for your codebase knowledge!",
-                        "inputSchema": {
-                            "type": "object",
-                            "properties": {
-                                "index_name": {
-                                    "type": "string",
-                                    "description": "Optional: Name of specific index to check. If not provided, shows status of all indexes.",
-                                }
-                            },
-                        },
-                    },
-                    {
-                        "name": "leann_list",
-                        "description": "📋 Show all your indexed codebases - your personal code library! Use this to see what's available for search.",
-                        "inputSchema": {"type": "object", "properties": {}},
-                    },
-                ]
-            },
-        }
-
-    elif request.get("method") == "tools/call":
-        tool_name = request["params"]["name"]
-        args = request["params"].get("arguments", {})
-
-        try:
-            if tool_name == "leann_search":
-                # Validate required parameters
-                if not args.get("index_name") or not args.get("query"):
-                    return {
-                        "jsonrpc": "2.0",
-                        "id": request.get("id"),
-                        "result": {
-                            "content": [
-                                {
-                                    "type": "text",
-                                    "text": "Error: Both index_name and query are required",
-                                }
-                            ]
-                        },
-                    }
-
-                # Build simplified command
-                cmd = [
-                    "leann",
-                    "search",
-                    args["index_name"],
-                    args["query"],
-                    f"--top-k={args.get('top_k', 5)}",
-                    f"--complexity={args.get('complexity', 32)}",
-                ]
-                result = subprocess.run(cmd, capture_output=True, text=True)
-
-            elif tool_name == "leann_status":
-                if args.get("index_name"):
-                    # Check specific index status - for now, we'll use leann list and filter
-                    result = subprocess.run(["leann", "list"], capture_output=True, text=True)
-                    # We could enhance this to show more detailed status per index
-                else:
-                    # Show all indexes status
-                    result = subprocess.run(["leann", "list"], capture_output=True, text=True)
-
-            elif tool_name == "leann_list":
-                result = subprocess.run(["leann", "list"], capture_output=True, text=True)
-
-            return {
-                "jsonrpc": "2.0",
-                "id": request.get("id"),
-                "result": {
-                    "content": [
-                        {
-                            "type": "text",
-                            "text": result.stdout
-                            if result.returncode == 0
-                            else f"Error: {result.stderr}",
-                        }
-                    ]
-                },
-            }
-
-        except Exception as e:
-            return {
-                "jsonrpc": "2.0",
-                "id": request.get("id"),
-                "error": {"code": -1, "message": str(e)},
-            }
-
-
-def main():
-    for line in sys.stdin:
-        try:
-            request = json.loads(line.strip())
-            response = handle_request(request)
-            if response:
-                print(json.dumps(response))
-                sys.stdout.flush()
-        except Exception as e:
-            error_response = {
-                "jsonrpc": "2.0",
-                "id": None,
-                "error": {"code": -1, "message": str(e)},
-            }
-            print(json.dumps(error_response))
-            sys.stdout.flush()
-
-
-if __name__ == "__main__":
-    main()

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

@@ -1,7 +1,7 @@
 import json
 from abc import ABC, abstractmethod
 from pathlib import Path
-from typing import Any, Literal, Optional
+from typing import Any, Literal
 
 import numpy as np
 
@@ -169,7 +169,7 @@ class BaseSearcher(LeannBackendSearcherInterface, ABC):
         prune_ratio: float = 0.0,
         recompute_embeddings: bool = False,
         pruning_strategy: Literal["global", "local", "proportional"] = "global",
-        zmq_port: Optional[int] = None,
+        zmq_port: int | None = None,
         **kwargs,
     ) -> dict[str, Any]:
         """

packages/leann-mcp/README.md

@@ -1,119 +0,0 @@
-# 🔥 LEANN Claude Code Integration
-
-Transform your development workflow with intelligent code assistance using LEANN's semantic search directly in Claude Code.
-
-## Prerequisites
-
-Install LEANN globally for MCP integration (with default backend):
-
-```bash
-uv tool install leann-core --with leann
-```
-This installs the `leann` CLI into an isolated tool environment and includes both backends so `leann build` works out-of-the-box.
-
-## 🚀 Quick Setup
-
-Add the LEANN MCP server to Claude Code:
-
-```bash
-claude mcp add leann-server -- leann_mcp
-```
-
-## 🛠️ Available Tools
-
-Once connected, you'll have access to these powerful semantic search tools in Claude Code:
-
-- **`leann_list`** - List all available indexes across your projects
-- **`leann_search`** - Perform semantic searches across code and documents
-- **`leann_ask`** - Ask natural language questions and get AI-powered answers from your codebase
-
-## 🎯 Quick Start Example
-
-```bash
-# Build an index for your project (change to your actual path)
-leann build my-project --docs ./
-
-# Start Claude Code
-claude
-```
-
-## 🚀 Advanced Usage Examples
-
-### Index Entire Git Repository
-```bash
-# Index all tracked files in your git repository, note right now we will skip submodules, but we can add it back easily if you want
-leann build my-repo --docs $(git ls-files) --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-
-# Index only specific file types from git
-leann build my-python-code --docs $(git ls-files "*.py") --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-```
-
-### Multiple Directories and Files
-```bash
-# Index multiple directories
-leann build my-codebase --docs ./src ./tests ./docs ./config --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-
-# Mix files and directories
-leann build my-project --docs ./README.md ./src/ ./package.json ./docs/ --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-
-# Specific files only
-leann build my-configs --docs ./tsconfig.json ./package.json ./webpack.config.js --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-```
-
-### Advanced Git Integration
-```bash
-# Index recently modified files
-leann build recent-changes --docs $(git diff --name-only HEAD~10..HEAD) --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-
-# Index files matching pattern
-leann build frontend --docs $(git ls-files "*.tsx" "*.ts" "*.jsx" "*.js") --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-
-# Index documentation and config files
-leann build docs-and-configs --docs $(git ls-files "*.md" "*.yml" "*.yaml" "*.json" "*.toml") --embedding-mode sentence-transformers --embedding-model all-MiniLM-L6-v2 --backend hnsw
-```
-
-
-**Try this in Claude Code:**
-```
-Help me understand this codebase. List available indexes and search for authentication patterns.
-```
-
-<p align="center">
-  <img src="../../assets/claude_code_leann.png" alt="LEANN in Claude Code" width="80%">
-</p>
-
-
-## 🧠 How It Works
-
-The integration consists of three key components working seamlessly together:
-
-- **`leann`** - Core CLI tool for indexing and searching (installed globally via `uv tool install`)
-- **`leann_mcp`** - MCP server that wraps `leann` commands for Claude Code integration
-- **Claude Code** - Calls `leann_mcp`, which executes `leann` commands and returns intelligent results
-
-## 📁 File Support
-
-LEANN understands **30+ file types** including:
-- **Programming**: Python, JavaScript, TypeScript, Java, Go, Rust, C++, C#
-- **Data**: SQL, YAML, JSON, CSV, XML
-- **Documentation**: Markdown, TXT, PDF
-- **And many more!**
-
-## 💾 Storage & Organization
-
-- **Project indexes**: Stored in `.leann/` directory (just like `.git`)
-- **Global registry**: Project tracking at `~/.leann/projects.json`
-- **Multi-project support**: Switch between different codebases seamlessly
-- **Portable**: Transfer indexes between machines with minimal overhead
-
-## 🗑️ Uninstalling
-
-To remove the LEANN MCP server from Claude Code:
-
-```bash
-claude mcp remove leann-server
-```
-To remove LEANN
-```
-uv pip uninstall leann leann-backend-hnsw leann-core
-```

packages/leann/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "leann"
-version = "0.2.9"
+version = "0.2.1"
 description = "LEANN - The smallest vector index in the world. RAG Everything with LEANN!"
 readme = "README.md"
 requires-python = ">=3.9"

pyproject.toml

@@ -32,7 +32,7 @@ dependencies = [
     "pypdfium2>=4.30.0",
     # LlamaIndex core and readers - updated versions
     "llama-index>=0.12.44",
-    "llama-index-readers-file>=0.4.0", # Essential for PDF parsing
+    "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",
@@ -40,13 +40,9 @@ dependencies = [
     # Other dependencies
     "ipykernel==6.29.5",
     "msgpack>=1.1.1",
-    "mlx>=0.26.3; sys_platform == 'darwin' and platform_machine == 'arm64'",
+    "mlx>=0.26.3; sys_platform == 'darwin'",
-    "mlx-lm>=0.26.0; sys_platform == 'darwin' and platform_machine == 'arm64'",
+    "mlx-lm>=0.26.0; sys_platform == 'darwin'",
     "psutil>=5.8.0",
-    "pybind11>=3.0.0",
-    "pathspec>=0.12.1",
-    "nbconvert>=7.16.6",
-    "gitignore-parser>=0.1.12",
 ]
 
 [project.optional-dependencies]
@@ -55,7 +51,7 @@ dev = [
     "pytest-cov>=4.0",
     "pytest-xdist>=3.0",  # For parallel test execution
     "black>=23.0",
-    "ruff==0.12.7",  # Fixed version to ensure consistent formatting across all environments
+    "ruff>=0.1.0",
     "matplotlib",
     "huggingface-hub>=0.20.0",
     "pre-commit>=3.5.0",
@@ -92,7 +88,7 @@ leann-backend-diskann = { path = "packages/leann-backend-diskann", editable = tr
 leann-backend-hnsw = { path = "packages/leann-backend-hnsw", editable = true }
 
 [tool.ruff]
-target-version = "py39"
+target-version = "py310"
 line-length = 100
 extend-exclude = [
     "third_party",
@@ -155,7 +151,7 @@ markers = [
     "slow: marks tests as slow (deselect with '-m \"not slow\"')",
     "openai: marks tests that require OpenAI API key",
 ]
-timeout = 300  # Reduced from 600s (10min) to 300s (5min) for CI safety
+timeout = 600
 addopts = [
     "-v",
     "--tb=short",

sky/leann-build.yaml

@@ -1,76 +0,0 @@
-name: leann-build
-
-resources:
-  # Choose a GPU for fast embeddings (examples: L4, A10G, A100). CPU also works but is slower.
-  accelerators: L4:1
-  # Optionally pin a cloud, otherwise SkyPilot will auto-select
-  # cloud: aws
-  disk_size: 100
-
-envs:
-  # Build parameters (override with: sky launch -c leann-gpu sky/leann-build.yaml -e key=value)
-  index_name: my-index
-  docs: ./data
-  backend: hnsw               # hnsw | diskann
-  complexity: 64
-  graph_degree: 32
-  num_threads: 8
-  # Embedding selection
-  embedding_mode: sentence-transformers   # sentence-transformers | openai | mlx | ollama
-  embedding_model: facebook/contriever
-  # Storage/latency knobs
-  recompute: true             # true => selective recomputation (recommended)
-  compact: true               # for HNSW only
-  # Optional pass-through
-  extra_args: ""
-  # Rebuild control
-  force: true
-
-# Sync local paths to the remote VM. Adjust as needed.
-file_mounts:
-  # Example: mount your local data directory used for building
-  ~/leann-data: ${docs}
-
-setup: |
-  set -e
-  # Install uv (package manager)
-  curl -LsSf https://astral.sh/uv/install.sh | sh
-  export PATH="$HOME/.local/bin:$PATH"
-
-  # Ensure modern libstdc++ for FAISS (GLIBCXX >= 3.4.30)
-  sudo apt-get update -y
-  sudo apt-get install -y libstdc++6 libgomp1
-  # Also upgrade conda's libstdc++ in base env (Skypilot images include conda)
-  if command -v conda >/dev/null 2>&1; then
-    conda install -y -n base -c conda-forge libstdcxx-ng
-  fi
-
-  # Install LEANN CLI and backends into the user environment
-  uv pip install --upgrade pip
-  uv pip install leann-core leann-backend-hnsw leann-backend-diskann
-
-run: |
-  export PATH="$HOME/.local/bin:$PATH"
-  # Derive flags from env
-  recompute_flag=""
-  if [ "${recompute}" = "false" ] || [ "${recompute}" = "0" ]; then
-    recompute_flag="--no-recompute"
-  fi
-  force_flag=""
-  if [ "${force}" = "true" ] || [ "${force}" = "1" ]; then
-    force_flag="--force"
-  fi
-
-  # Build command
-  python -m leann.cli build ${index_name} \
-    --docs ~/leann-data \
-    --backend ${backend} \
-    --complexity ${complexity} \
-    --graph-degree ${graph_degree} \
-    --num-threads ${num_threads} \
-    --embedding-mode ${embedding_mode} \
-    --embedding-model ${embedding_model} \
-    ${recompute_flag} ${force_flag} ${extra_args}
-
-  # Print where the index is stored for downstream rsync
-  echo "INDEX_OUT_DIR=~/.leann/indexes/${index_name}"

tests/README.md

@@ -6,11 +6,10 @@ This directory contains automated tests for the LEANN project using pytest.
 
 ### `test_readme_examples.py`
 Tests the examples shown in README.md:
-- The basic example code that users see first (parametrized for both HNSW and DiskANN backends)
+- The basic example code that users see first
 - Import statements work correctly
 - Different backend options (HNSW, DiskANN)
-- Different LLM configuration options (parametrized for both backends)
+- Different LLM configuration options
-- **All main README examples are tested with both HNSW and DiskANN backends using pytest parametrization**
 
 ### `test_basic.py`
 Basic functionality tests that verify:
@@ -26,16 +25,6 @@ Tests the document RAG example functionality:
 - Tests error handling with invalid parameters
 - Verifies that normalized embeddings are detected and cosine distance is used
 
-### `test_diskann_partition.py`
-Tests DiskANN graph partitioning functionality:
-- Tests DiskANN index building without partitioning (baseline)
-- Tests automatic graph partitioning with `is_recompute=True`
-- Verifies that partition files are created and large files are cleaned up for storage saving
-- Tests search functionality with partitioned indices
-- Validates medoid and max_base_norm file generation and usage
-- Includes performance comparison between DiskANN (with partition) and HNSW
-- **Note**: These tests are skipped in CI due to hardware requirements and computation time
-
 ## Running Tests
 
 ### Install test dependencies:
@@ -65,23 +54,15 @@ pytest tests/ -m "not openai"
 
 # Skip slow tests
 pytest tests/ -m "not slow"
-
-# Run DiskANN partition tests (requires local machine, not CI)
-pytest tests/test_diskann_partition.py
 ```
 
 ### Run with specific backend:
 ```bash
 # Test only HNSW backend
 pytest tests/test_basic.py::test_backend_basic[hnsw]
-pytest tests/test_readme_examples.py::test_readme_basic_example[hnsw]
 
 # Test only DiskANN backend
 pytest tests/test_basic.py::test_backend_basic[diskann]
-pytest tests/test_readme_examples.py::test_readme_basic_example[diskann]
-
-# All DiskANN tests (parametrized + specialized partition tests)
-pytest tests/ -k diskann
 ```
 
 ## CI/CD Integration

tests/test_basic.py

@@ -64,9 +64,6 @@ def test_backend_basic(backend_name):
         assert isinstance(results[0], SearchResult)
         assert "topic 2" in results[0].text or "document" in results[0].text
 
-        # Ensure cleanup to avoid hanging background servers
-        searcher.cleanup()
-
 
 @pytest.mark.skipif(
     os.environ.get("CI") == "true", reason="Skip model tests in CI to avoid MPS memory issues"
@@ -93,5 +90,3 @@ def test_large_index():
         searcher = LeannSearcher(index_path)
         results = searcher.search(["word10 word20"], top_k=10)
         assert len(results[0]) == 10
-        # Cleanup
-        searcher.cleanup()

tests/test_diskann_partition.py

@@ -1,369 +0,0 @@
-"""
-Test DiskANN graph partitioning functionality.
-
-Tests the automatic graph partitioning feature that was implemented to save
-storage space by partitioning large DiskANN indices and safely deleting
-redundant files while maintaining search functionality.
-"""
-
-import os
-import tempfile
-from pathlib import Path
-
-import pytest
-
-
-@pytest.mark.skipif(
-    os.environ.get("CI") == "true",
-    reason="Skip DiskANN partition tests in CI - requires specific hardware and large memory",
-)
-def test_diskann_without_partition():
-    """Test DiskANN index building without partition (baseline)."""
-    from leann.api import LeannBuilder, LeannSearcher
-
-    with tempfile.TemporaryDirectory() as temp_dir:
-        index_path = str(Path(temp_dir) / "test_no_partition.leann")
-
-        # Test data - enough to trigger index building
-        texts = [
-            f"Document {i} discusses topic {i % 10} with detailed analysis of subject {i // 10}."
-            for i in range(500)
-        ]
-
-        # Build without partition (is_recompute=False)
-        builder = LeannBuilder(
-            backend_name="diskann",
-            embedding_model="facebook/contriever",
-            embedding_mode="sentence-transformers",
-            num_neighbors=32,
-            search_list_size=50,
-            is_recompute=False,  # No partition
-        )
-
-        for text in texts:
-            builder.add_text(text)
-
-        builder.build_index(index_path)
-
-        # Verify index was created
-        index_dir = Path(index_path).parent
-        assert index_dir.exists()
-
-        # Check that traditional DiskANN files exist
-        index_prefix = Path(index_path).stem
-        # Core DiskANN files (beam search index may not be created for small datasets)
-        required_files = [
-            f"{index_prefix}_disk.index",
-            f"{index_prefix}_pq_compressed.bin",
-            f"{index_prefix}_pq_pivots.bin",
-        ]
-
-        # Check all generated files first for debugging
-        generated_files = [f.name for f in index_dir.glob(f"{index_prefix}*")]
-        print(f"Generated files: {generated_files}")
-
-        for required_file in required_files:
-            file_path = index_dir / required_file
-            assert file_path.exists(), f"Required file {required_file} not found"
-
-        # Ensure no partition files exist in non-partition mode
-        partition_files = [f"{index_prefix}_disk_graph.index", f"{index_prefix}_partition.bin"]
-
-        for partition_file in partition_files:
-            file_path = index_dir / partition_file
-            assert not file_path.exists(), (
-                f"Partition file {partition_file} should not exist in non-partition mode"
-            )
-
-        # Test search functionality
-        searcher = LeannSearcher(index_path)
-        results = searcher.search("topic 3 analysis", top_k=3)
-
-        assert len(results) > 0
-        assert all(result.score is not None and result.score != float("-inf") for result in results)
-
-
-@pytest.mark.skipif(
-    os.environ.get("CI") == "true",
-    reason="Skip DiskANN partition tests in CI - requires specific hardware and large memory",
-)
-def test_diskann_with_partition():
-    """Test DiskANN index building with automatic graph partitioning."""
-    from leann.api import LeannBuilder
-
-    with tempfile.TemporaryDirectory() as temp_dir:
-        index_path = str(Path(temp_dir) / "test_with_partition.leann")
-
-        # Test data - enough to trigger partitioning
-        texts = [
-            f"Document {i} explores subject {i % 15} with comprehensive coverage of area {i // 15}."
-            for i in range(500)
-        ]
-
-        # Build with partition (is_recompute=True)
-        builder = LeannBuilder(
-            backend_name="diskann",
-            embedding_model="facebook/contriever",
-            embedding_mode="sentence-transformers",
-            num_neighbors=32,
-            search_list_size=50,
-            is_recompute=True,  # Enable automatic partitioning
-        )
-
-        for text in texts:
-            builder.add_text(text)
-
-        builder.build_index(index_path)
-
-        # Verify index was created
-        index_dir = Path(index_path).parent
-        assert index_dir.exists()
-
-        # Check that partition files exist
-        index_prefix = Path(index_path).stem
-        partition_files = [
-            f"{index_prefix}_disk_graph.index",  # Partitioned graph
-            f"{index_prefix}_partition.bin",  # Partition metadata
-            f"{index_prefix}_pq_compressed.bin",
-            f"{index_prefix}_pq_pivots.bin",
-        ]
-
-        for partition_file in partition_files:
-            file_path = index_dir / partition_file
-            assert file_path.exists(), f"Expected partition file {partition_file} not found"
-
-        # Check that large files were cleaned up (storage saving goal)
-        large_files = [f"{index_prefix}_disk.index", f"{index_prefix}_disk_beam_search.index"]
-
-        for large_file in large_files:
-            file_path = index_dir / large_file
-            assert not file_path.exists(), (
-                f"Large file {large_file} should have been deleted for storage saving"
-            )
-
-        # Verify required auxiliary files for partition mode exist
-        required_files = [
-            f"{index_prefix}_disk.index_medoids.bin",
-            f"{index_prefix}_disk.index_max_base_norm.bin",
-        ]
-
-        for req_file in required_files:
-            file_path = index_dir / req_file
-            assert file_path.exists(), (
-                f"Required auxiliary file {req_file} missing for partition mode"
-            )
-
-
-@pytest.mark.skipif(
-    os.environ.get("CI") == "true",
-    reason="Skip DiskANN partition tests in CI - requires specific hardware and large memory",
-)
-def test_diskann_partition_search_functionality():
-    """Test that search works correctly with partitioned indices."""
-    from leann.api import LeannBuilder, LeannSearcher
-
-    with tempfile.TemporaryDirectory() as temp_dir:
-        index_path = str(Path(temp_dir) / "test_partition_search.leann")
-
-        # Create diverse test data
-        texts = [
-            "LEANN is a storage-efficient approximate nearest neighbor search system.",
-            "Graph partitioning helps reduce memory usage in large scale vector search.",
-            "DiskANN provides high-performance disk-based approximate nearest neighbor search.",
-            "Vector embeddings enable semantic search over unstructured text data.",
-            "Approximate nearest neighbor algorithms trade accuracy for speed and storage.",
-        ] * 100  # Repeat to get enough data
-
-        # Build with partitioning
-        builder = LeannBuilder(
-            backend_name="diskann",
-            embedding_model="facebook/contriever",
-            embedding_mode="sentence-transformers",
-            is_recompute=True,  # Enable partitioning
-        )
-
-        for text in texts:
-            builder.add_text(text)
-
-        builder.build_index(index_path)
-
-        # Test search with partitioned index
-        searcher = LeannSearcher(index_path)
-
-        # Test various queries
-        test_queries = [
-            ("vector search algorithms", 5),
-            ("LEANN storage efficiency", 3),
-            ("graph partitioning memory", 4),
-            ("approximate nearest neighbor", 7),
-        ]
-
-        for query, top_k in test_queries:
-            results = searcher.search(query, top_k=top_k)
-
-            # Verify search results
-            assert len(results) == top_k, f"Expected {top_k} results for query '{query}'"
-            assert all(result.score is not None for result in results), (
-                "All results should have scores"
-            )
-            assert all(result.score != float("-inf") for result in results), (
-                "No result should have -inf score"
-            )
-            assert all(result.text is not None for result in results), (
-                "All results should have text"
-            )
-
-            # Scores should be in descending order (higher similarity first)
-            scores = [result.score for result in results]
-            assert scores == sorted(scores, reverse=True), (
-                "Results should be sorted by score descending"
-            )
-
-
-@pytest.mark.skipif(
-    os.environ.get("CI") == "true",
-    reason="Skip DiskANN partition tests in CI - requires specific hardware and large memory",
-)
-def test_diskann_medoid_and_norm_files():
-    """Test that medoid and max_base_norm files are correctly generated and used."""
-    import struct
-
-    from leann.api import LeannBuilder, LeannSearcher
-
-    with tempfile.TemporaryDirectory() as temp_dir:
-        index_path = str(Path(temp_dir) / "test_medoid_norm.leann")
-
-        # Small but sufficient dataset
-        texts = [f"Test document {i} with content about subject {i % 10}." for i in range(200)]
-
-        builder = LeannBuilder(
-            backend_name="diskann",
-            embedding_model="facebook/contriever",
-            embedding_mode="sentence-transformers",
-            is_recompute=True,
-        )
-
-        for text in texts:
-            builder.add_text(text)
-
-        builder.build_index(index_path)
-
-        index_dir = Path(index_path).parent
-        index_prefix = Path(index_path).stem
-
-        # Test medoids file
-        medoids_file = index_dir / f"{index_prefix}_disk.index_medoids.bin"
-        assert medoids_file.exists(), "Medoids file should be generated"
-
-        # Read and validate medoids file format
-        with open(medoids_file, "rb") as f:
-            nshards = struct.unpack("<I", f.read(4))[0]
-            one_val = struct.unpack("<I", f.read(4))[0]
-            medoid_id = struct.unpack("<I", f.read(4))[0]
-
-            assert nshards == 1, "Single-shot build should have 1 shard"
-            assert one_val == 1, "Expected value should be 1"
-            assert medoid_id >= 0, "Medoid ID should be valid (not hardcoded 0)"
-
-        # Test max_base_norm file
-        norm_file = index_dir / f"{index_prefix}_disk.index_max_base_norm.bin"
-        assert norm_file.exists(), "Max base norm file should be generated"
-
-        # Read and validate norm file
-        with open(norm_file, "rb") as f:
-            npts = struct.unpack("<I", f.read(4))[0]
-            ndims = struct.unpack("<I", f.read(4))[0]
-            norm_val = struct.unpack("<f", f.read(4))[0]
-
-            assert npts == 1, "Should have 1 norm point"
-            assert ndims == 1, "Should have 1 dimension"
-            assert norm_val > 0, "Norm value should be positive"
-            assert norm_val != float("inf"), "Norm value should be finite"
-
-        # Test that search works with these files
-        searcher = LeannSearcher(index_path)
-        results = searcher.search("test subject", top_k=3)
-
-        # Verify that scores are not -inf (which indicates norm file was loaded correctly)
-        assert len(results) > 0
-        assert all(result.score != float("-inf") for result in results), (
-            "Scores should not be -inf when norm file is correct"
-        )
-
-
-@pytest.mark.skipif(
-    os.environ.get("CI") == "true",
-    reason="Skip performance comparison in CI - requires significant compute time",
-)
-def test_diskann_vs_hnsw_performance():
-    """Compare DiskANN (with partition) vs HNSW performance."""
-    import time
-
-    from leann.api import LeannBuilder, LeannSearcher
-
-    with tempfile.TemporaryDirectory() as temp_dir:
-        # Test data
-        texts = [
-            f"Performance test document {i} covering topic {i % 20} in detail." for i in range(1000)
-        ]
-        query = "performance topic test"
-
-        # Test DiskANN with partitioning
-        diskann_path = str(Path(temp_dir) / "perf_diskann.leann")
-        diskann_builder = LeannBuilder(
-            backend_name="diskann",
-            embedding_model="facebook/contriever",
-            embedding_mode="sentence-transformers",
-            is_recompute=True,
-        )
-
-        for text in texts:
-            diskann_builder.add_text(text)
-
-        start_time = time.time()
-        diskann_builder.build_index(diskann_path)
-
-        # Test HNSW
-        hnsw_path = str(Path(temp_dir) / "perf_hnsw.leann")
-        hnsw_builder = LeannBuilder(
-            backend_name="hnsw",
-            embedding_model="facebook/contriever",
-            embedding_mode="sentence-transformers",
-            is_recompute=True,
-        )
-
-        for text in texts:
-            hnsw_builder.add_text(text)
-
-        start_time = time.time()
-        hnsw_builder.build_index(hnsw_path)
-
-        # Compare search performance
-        diskann_searcher = LeannSearcher(diskann_path)
-        hnsw_searcher = LeannSearcher(hnsw_path)
-
-        # Warm up searches
-        diskann_searcher.search(query, top_k=5)
-        hnsw_searcher.search(query, top_k=5)
-
-        # Timed searches
-        start_time = time.time()
-        diskann_results = diskann_searcher.search(query, top_k=10)
-        diskann_search_time = time.time() - start_time
-
-        start_time = time.time()
-        hnsw_results = hnsw_searcher.search(query, top_k=10)
-        hnsw_search_time = time.time() - start_time
-
-        # Basic assertions
-        assert len(diskann_results) == 10
-        assert len(hnsw_results) == 10
-        assert all(r.score != float("-inf") for r in diskann_results)
-        assert all(r.score != float("-inf") for r in hnsw_results)
-
-        # Performance ratio (informational)
-        if hnsw_search_time > 0:
-            speed_ratio = hnsw_search_time / diskann_search_time
-            print(f"DiskANN search time: {diskann_search_time:.4f}s")
-            print(f"HNSW search time: {hnsw_search_time:.4f}s")
-            print(f"DiskANN is {speed_ratio:.2f}x faster than HNSW")

tests/test_document_rag.py

@@ -58,9 +58,6 @@ def test_document_rag_simulated(test_data_dir):
 
 
 @pytest.mark.skipif(not os.environ.get("OPENAI_API_KEY"), reason="OpenAI API key not available")
-@pytest.mark.skipif(
-    os.environ.get("CI") == "true", reason="Skip OpenAI tests in CI to avoid API costs"
-)
 def test_document_rag_openai(test_data_dir):
     """Test document_rag with OpenAI embeddings."""
     with tempfile.TemporaryDirectory() as temp_dir:

tests/test_readme_examples.py

@@ -10,33 +10,29 @@ from pathlib import Path
 import pytest
 
 
-@pytest.mark.parametrize("backend_name", ["hnsw", "diskann"])
+def test_readme_basic_example():
-def test_readme_basic_example(backend_name):
+    """Test the basic example from README.md."""
-    """Test the basic example from README.md with both backends."""
     # Skip on macOS CI due to MPS environment issues with all-MiniLM-L6-v2
     if os.environ.get("CI") == "true" and platform.system() == "Darwin":
         pytest.skip("Skipping on macOS CI due to MPS environment issues with all-MiniLM-L6-v2")
-    # Skip DiskANN on CI (Linux runners) due to C++ extension memory/hardware constraints
-    if os.environ.get("CI") == "true" and backend_name == "diskann":
-        pytest.skip("Skip DiskANN tests in CI due to resource constraints and instability")
 
     # This is the exact code from README (with smaller model for CI)
     from leann import LeannBuilder, LeannChat, LeannSearcher
     from leann.api import SearchResult
 
     with tempfile.TemporaryDirectory() as temp_dir:
-        INDEX_PATH = str(Path(temp_dir) / f"demo_{backend_name}.leann")
+        INDEX_PATH = str(Path(temp_dir) / "demo.leann")
 
         # Build an index
         # In CI, use a smaller model to avoid memory issues
         if os.environ.get("CI") == "true":
             builder = LeannBuilder(
-                backend_name=backend_name,
+                backend_name="hnsw",
                 embedding_model="sentence-transformers/all-MiniLM-L6-v2",  # Smaller model
                 dimensions=384,  # Smaller dimensions
             )
         else:
-            builder = LeannBuilder(backend_name=backend_name)
+            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)
@@ -56,15 +52,9 @@ def test_readme_basic_example(backend_name):
         # Verify search results
         assert len(results) > 0
         assert isinstance(results[0], SearchResult)
-        assert results[0].score != float("-inf"), (
-            f"should return valid scores, got {results[0].score}"
-        )
         # The second text about banana-crocodile should be more relevant
         assert "banana" in results[0].text or "crocodile" in results[0].text
 
-        # Ensure we cleanup background embedding server
-        searcher.cleanup()
-
         # Chat with your data (using simulated LLM to avoid external dependencies)
         chat = LeannChat(INDEX_PATH, llm_config={"type": "simulated"})
         response = chat.ask("How much storage does LEANN save?", top_k=1)
@@ -72,8 +62,6 @@ def test_readme_basic_example(backend_name):
         # Verify chat works
         assert isinstance(response, str)
         assert len(response) > 0
-        # Cleanup chat resources
-        chat.cleanup()
 
 
 def test_readme_imports():
@@ -122,31 +110,26 @@ def test_backend_options():
         assert len(list(Path(diskann_path).parent.glob(f"{Path(diskann_path).stem}.*"))) > 0
 
 
-@pytest.mark.parametrize("backend_name", ["hnsw", "diskann"])
+def test_llm_config_simulated():
-def test_llm_config_simulated(backend_name):
+    """Test simulated LLM configuration option."""
-    """Test simulated LLM configuration option with both backends."""
     # Skip on macOS CI due to MPS environment issues with all-MiniLM-L6-v2
     if os.environ.get("CI") == "true" and platform.system() == "Darwin":
         pytest.skip("Skipping on macOS CI due to MPS environment issues with all-MiniLM-L6-v2")
 
-    # Skip DiskANN tests in CI due to hardware requirements
-    if os.environ.get("CI") == "true" and backend_name == "diskann":
-        pytest.skip("Skip DiskANN tests in CI - requires specific hardware and large memory")
-
     from leann import LeannBuilder, LeannChat
 
     with tempfile.TemporaryDirectory() as temp_dir:
         # Build a simple index
-        index_path = str(Path(temp_dir) / f"test_{backend_name}.leann")
+        index_path = str(Path(temp_dir) / "test.leann")
         # Use smaller model in CI to avoid memory issues
         if os.environ.get("CI") == "true":
             builder = LeannBuilder(
-                backend_name=backend_name,
+                backend_name="hnsw",
                 embedding_model="sentence-transformers/all-MiniLM-L6-v2",
                 dimensions=384,
             )
         else:
-            builder = LeannBuilder(backend_name=backend_name)
+            builder = LeannBuilder(backend_name="hnsw")
         builder.add_text("Test document for LLM testing")
         builder.build_index(index_path)