pgvector
pgvector is an open-source Postgres extension for vector similarity search. Vectors live in ordinary Postgres tables beside relational data, so one database holds both. Postgres 13 or later is supported.
Repository snapshot (GitHub API, 2026-11-19)
| Item | Detail |
|---|---|
| Repo | pgvector/pgvector, C |
| Licence | GitHub reports no classification; the LICENSE text is a permissive grant with a PostgreSQL Global Development Group copyright notice, which matches the PostgreSQL License (inferred, not confirmed by GitHub) |
| Created | 2021-04-20 |
| Activity | last push 2026-11-19; newest git tag v0.8.7 (the repository publishes no GitHub releases) |
| Size | 23,734 stars, 1,401 forks |
Capabilities (README)
- Indexes: HNSW and IVFFlat (approximate), or exact search with no index.
- Types:
vector(single precision),halfvec,bitandsparsevec. - Distances: L2, inner product, cosine, L1, plus Hamming and Jaccard for binary vectors.
- Filtering: ordinary SQL
WHEREwith Postgres indexes. Filtering runs after an approximate index scan, so results can be missing; iterative index scans (from v0.8.0) mitigate this, and partial indexes and partitioning are recommended. - Stated limits: indexes cover
vectorup to 2,000 dimensions,halfvecup to 4,000,bitup to 64,000, andsparsevecup to 1,000 non-zero elements. Quantisation is suggested for scale. - Hybrid search is not presented as a feature; it is done with SQL plus Postgres full-text search.
Deployment and billing
An extension you install (source, Docker, Homebrew, APT, Yum, conda-forge and others) or get from hosted Postgres providers; there is no pgvector-specific billing model, and cost follows the Postgres host. The README names no company maintainer.
Where it fits
pgvector is the “keep it in Postgres” option: no new system, SQL joins, transactions. See postgres-for-ai-agents and vector-databases-compared; compare qdrant and milvus for dedicated engines.
Sources
- The pgvector README, LICENSE file, tags and GitHub API (2026-11-19)