Arroy is a Rust library for Approximate Nearest Neighbors (ANN) search, designed for efficient memory usage and concurrent access. It targets developers needing to find similar vectors in high-dimensional spaces, particularly those already using or considering LMDB for data storage, offering a low memory footprint and shared data access benefits.

How It Works

Arroy builds upon the Annoy library's random projection tree approach but utilizes LMDB as its backend storage. This memory-mapped key-value store allows multiple processes to share the same index atomically, enabling concurrent reads and writes. The library supports Euclidean, Manhattan, cosine, and dot product distances, with index creation being a separate, potentially multi-threaded process from querying.

Quick Start & Requirements

• Install via cargo install arroy.
• Requires Rust toolchain.
• See crates.io for usage examples.

Highlighted Details

• LMDB backend enables shared memory access and atomic updates across processes.
• Supports Euclidean, Manhattan, cosine, and dot product distances.
• Offers features beyond Annoy, including query filtering and incremental updates.
• Designed for low memory usage, performing well up to ~1,000 dimensions.

Maintenance & Community

• Developed by Meilisearch contributors.
• See GitHub Actions for build status.
• Docs.rs provides API documentation.

Licensing & Compatibility

• MIT License.
• Compatible with commercial and closed-source applications.

Limitations & Caveats

Arroy lacks Python bindings and Hamming distance support. Performance may be slower than non-LMDB alternatives due to LMDB's log(n) lookups and non-aligned vectors. Index creation is a distinct step, and items cannot be added after the tree is built.