PostgreSQL for AI. Real Vector Search in PostgreSQL
summary
PostgreSQL plus PGVector now hits production-scale vector search: benchmarks report 6–16× speedups and ~75% lower cost versus 2023 results, and a 2025 test on 100M vectors reaches ~10,000 q/s on 100 CPUs with INF + bit quantization. PGVector exposes HNSW and flat indexes with tunable M and EF affecting build time, index size and recall. Reducing floats to float2 halves storage; 1-bit quantization cuts index size ~20× and, with a two‑step search, keeps reasonable accuracy for >1k‑dim vectors. Filters break naive NN; use partial indexes, streaming scans or multicol vector index for correlated/selective predicates.
AI-written summary · about
chapters (10)
- 0:00Introduction to talk
- 1:04PostgreSQL and PG Vector growth
- 3:38What are vectors in AI
- 8:20PG Vector indexes overview
- 13:36Vector search in PostgreSQL
- 16:58Vector size reduction techniques
- 23:46Filtering with vector queries
- 28:27Solutions for low selectivity
- 32:47PostgreSQL advantages for vectors
- 34:34Q&A session
related talks
52mVector search is now boring, but PostgreSQL has ways to go
48mFrom Text to Meaning: pgvector Transforms PostgreSQL Search
53mVectors: how to better support a nasty data type in PostgreSQL
04mSupercharging GenAI PostgreSQL Vector Search Optimization
25mVector data in Postgres - how's it different from "normal" data?
52mpgvector Newbie to Expert
44mScaling Semantic Search in PostgreSQL
54mDissimilarity search...