vchord Documentation
vchord (VectorChord) is a PostgreSQL extension built on pgvector for scalable, high-performance, and disk-friendly vector similarity search. Powered by RaBitQ 1-bit/4-bit/8-bit quantization, it delivers inverted-file (vchordrq) and single-layer graph (vchordg) index access methods, native multi-vector MaxSim (@#), sphere range queries (<<->>), and external centroid precomputation.
Documentation Sections
Overview, quickstart tutorial, installation via Pigsty, Docker, or prebuilt packages, and migration from pgvecto.rs.
Indexing with vchordrq and vchordg, nearest-neighbor and sphere search, filtering, prewarm, external index builds, and full SQL reference.
Multi-vector ColBERT MaxSim (@#) late interaction and hybrid semantic + BM25 search with vchord_bm25 and pg_tokenizer.
Reproducible performance, recall, memory, and cost comparisons across GIST-1M, Cohere-10M, and LAION-100M.
Production parameter tuning, memory sizing, I/O concurrency, vacuuming, and version upgrades.
How the University of Oxford stores and queries 3 billion satellite embedding vectors in PostgreSQL with vchord.
Key Topics & Deep Dives
Compare disk-friendly hierarchical K-Means (vchordrq) against low-latency single-layer graphs (vchordg) and configure build options.
Compress 32-bit vectors into 4-bit (rabitq4) or 8-bit (rabitq8) quantized types with theoretical error bounds.
Index vector[] token embeddings with WARP dynamic pruning for ColBERTv2 and ColPali visual document retrieval.
Offload K-Means centroid clustering to external CPUs or GPUs and build 100M-vector indexes in 20 minutes.
Tune probes, epsilon, ef_search, io_search, max_scan_tuples, and PostgreSQL shared memory for production workloads.
Step-by-step SQL migration guide from pgvecto.rs (vectors extension) or pgvector HNSW/IVFFlat to vchord.