hnswlib-rs

Overview

hnswlib-rs is a pure-Rust implementation of the HNSW (Hierarchical Navigable Small World) algorithm inspired by the original C++ hnswlib. Designed with a decoupled architecture separating graph structure from vector storage.

Key Features

• Decoupled Architecture: Graph separated from vector storage for flexibility
• External Key Support: Map your external keys to internal NodeId
• Multiple Precision: f32, f16, bf16, and per-vector quantized int8
• Pure Rust: Memory-safe implementation without C dependencies
• Flexible Storage: Provide vectors on demand via VectorStore interface

Architecture

The library intentionally decouples components:

• Hnsw<K, M>: Owns the graph + mapping from external key K to internal NodeId
• VectorStore: Keyed by NodeId, supplies vectors on demand
• This design allows flexible storage backends and memory management

Supported Data Types

• f32: Standard 32-bit floating point
• f16: 16-bit floating point for memory efficiency
• bf16: Brain floating point for ML applications
• int8: Quantized 8-bit integers for maximum compression

Use Cases

• Approximate nearest neighbor search
• Embedding similarity search
• Large-scale vector retrieval
• Custom storage backend integration
• Applications requiring memory-efficient vector search

Installation

Available on crates.io:

[dependencies]
hnswlib-rs = "*"

Comparison to Other Implementations

vs. C++ hnswlib

• Memory safety guarantees from Rust
• No C dependencies or FFI overhead
• Modern Rust ecosystem integration

vs. Other Rust HNSW

• Unique decoupled architecture
• Flexible vector storage
• Support for multiple precision formats

Repository

GitHub: jean-pierreBoth/hnswlib-rs Crates.io: https://crates.io/crates/hnswlib-rs

Technical Details

Implements the algorithm from: "Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs" by Yu. A. Malkov and D. A. Yashunin

Pricing

Free and open-source under MIT or Apache 2.0 license. No licensing costs.