VectorChord
A PostgreSQL extension for scalable, disk-friendly vector search, the successor to pgvecto.rs, that compresses vectors with RaBitQ and autonomous reranking to cut storage cost while preserving search quality.
VectorChord is a PostgreSQL extension for scalable, high-performance and cost-effective vector search, and the successor to pgvecto.rs. It applies RaBitQ compression together with autonomous reranking to store vectors efficiently while preserving search quality, and the project reports storing 400,000 vectors for one dollar, several times more than pgvector or pgvecto.rs at the same price, and hosting 100 million 768-dimensional vectors on a single mid-sized cloud instance.
It reports indexing 100 million vectors in about 20 minutes using hierarchical K-means and optimized disk operations, and uses dimensionality reduction and sampling to control memory growth so that a billion-vector index can be built on a machine with 128GB of memory.
It is written in Rust and distributed as a Postgres extension, with Docker images available.