Merge code

CLAUDE.md

@@ -0,0 +1,79 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+PyGlass is a high-performance C++ library with Python bindings for approximate similarity search using graph-based algorithms (HNSW, NSG, SSG, RNNDESCENT, IVF). The core is implemented in C++ with SIMD optimizations, and Python bindings are provided via pybind11.
+
+## Essential Commands
+
+### Build and Installation
+```bash
+# Development installation (editable)
+pip install -v -e "python"
+
+# Using uv (recommended for development)
+uv venv
+source .venv/bin/activate
+uv pip install -v -e "python"
+
+# Generate C++ compile_commands.json
+bear -- uv pip install -v -e "python"
+```
+
+### Testing
+```bash
+pytest "python"
+```
+
+### Code Quality
+```bash
+# Python formatting
+black .
+
+# Python linting
+ruff check .
+
+# C++ formatting
+clang-format -i $(find glass -name "*.hpp")
+```
+
+### Benchmarking
+```bash
+# Run with default configuration
+python3 examples/main.py
+
+# Run with custom configuration
+python3 examples/main.py path/to/config.yaml
+```
+
+## Architecture
+
+### Core Components
+
+1. **Graph Algorithms** (`glass/hnsw/`, `glass/nsg/`): HNSW and NSG graph implementations for similarity search
+2. **Quantization** (`glass/quant/`): Multiple quantization methods (FP16, SQ8, PQ, etc.) for memory efficiency
+3. **SIMD Optimizations** (`glass/simd/`): Platform-specific optimizations for AVX2, AVX512, and ARM NEON
+4. **Search Layer** (`glass/searcher/`): Unified search interface over different index types
+
+### Python Integration
+
+- **Bindings**: `python/bindings.cc` uses pybind11 to expose C++ classes
+- **Builder Pattern**: Python API uses builder pattern for index construction
+- **Dataset Utils**: `python/ann_dataset/` handles standard ANN benchmark datasets
+
+### Key Design Patterns
+
+1. **Template-Heavy C++**: Extensive use of C++ templates for performance and flexibility
+2. **Memory Management**: Custom memory allocators and storage classes for efficient data handling
+3. **Parallel Processing**: OpenMP for multi-threaded operations
+4. **Builder Pattern**: Index construction uses fluent builder API
+
+## Development Notes
+
+- C++20 standard required
+- No external C++ dependencies (self-contained)
+- Performance is critical - always benchmark changes
+- Use `examples/config.yaml` to configure benchmark parameters
+- Dataset downloads may require `export HF_ENDPOINT=https://hf-mirror.com` for network issues

examples/config.yaml

@@ -6,24 +6,23 @@ index_params:
     index_args:
       R: 48
       L: 200
+      build_quant: SQ8U
 
 rebuild: false
-build_quant: SQ8U
+
 search_quants:
-  - SQ4U
-refine_quant: FP16
+  - search_quant: SQ4U
+    refine_quant: FP16
+
 topks:
-  - 10
+  - 100
+
 efs:
-  - 20
-  - 30
-  - 40
-  - 50
-  - 60
-  - 70
-  - 80
-  - 90
   - 100
-runs: 5
-batch:
-  - true
+  - 200
+  - 300
+
+runs: 3
+
+concurrency:
+  min: 1