Blaze-DB
A disk-backed vector database written in Rust that uses HNSW indexing and quantization for approximate nearest-neighbor search over embeddings.
Blaze-DB stores and indexes embeddings on disk in binary or JSON format and retrieves them using an HNSW (Hierarchical Navigable Small World) graph for approximate nearest-neighbor search, combined with quantization to reduce storage.
It ships as two binaries, a server and a client, for server and client operations, with batch and chunked processing for embedding generation in the CLI wrapper. Memory-mapped files and parallel processing support fast, concurrent, thread-safe reads and writes, and an LRU index cache gives substantially faster I/O.
Search supports multiple distance metrics, including cosine, Euclidean and dot product, and the project includes a performance benchmarking suite.