HNSWLIB

HNSWLIB is a header-only C++ library with Python bindings for fast approximate nearest neighbor (ANN) search using Hierarchical Navigable Small World (HNSW) graphs.

Features

• Fast Approximate Nearest Neighbor Search: Implements the HNSW algorithm for efficient high-dimensional vector search.
• Header-only C++ Implementation: No dependencies other than C++11.
• Python Bindings: Full support for Python with picklable index objects.
• Incremental Index Construction: Supports adding, updating, and deleting elements from the index.
• Multi-threading Support: Multi-threaded index construction and search; configurable number of threads.
• Custom Distances: C++ interface allows user-defined distance functions.
• Supported Distance Metrics:
  • Squared L2 ('l2')
  • Inner Product ('ip')
  • Cosine Similarity ('cosine')
• Filtering Support: Ability to filter search results by element labels.
• Efficient Memory Usage: Lower memory footprint and faster build time compared to some alternatives.
• Serialization/Deserialization: Save/load indexes to/from disk; supports pickle in Python.
• Replace Deleted Elements: Option to reuse memory of deleted elements.
• Index Resizing: Dynamically increase/shrink the index capacity.
• Multi-vector Document Search and Epsilon Search: Available in C++ interface.
• Java and R Support: External bindings available.
• Cross-platform: Usable in C++ and Python environments; external implementations for other languages (Go, Julia, Rust, Java, .NET, etc.).

Usage Highlights

• Create, initialize, and manage ANN indexes in both C++ and Python.
• Insert, update, delete, and query vectors.
• Batch operations for both insertions and queries.
• Save/load indexes for persistence and scaling.
• Fine-tune speed/accuracy trade-offs with parameters like ef and M.

Installation

• Install via pip: pip install hnswlib
• Or build from source (see GitHub for details).

License

• Apache-2.0

Pricing

• Open Source: Free to use under the Apache-2.0 license.

Resources

• GitHub Repository
• Documentation & Examples
• HNSW Paper