pgvector
Vector similarity search for PostgreSQL
Our Take
pgvector's maturation is the biggest shift of 2025. Version 0.8.0 brought iterative scans fixing the notorious overfiltering problem, plus improved cost estimation. The companion pgvectorscale extension adds DiskANN-based indexing with 9x smaller indexes than HNSW. On 50M Cohere embeddings, pgvectorscale achieves 471 QPS at 99% recall with 28ms p95 — beating Pinecone s1 by 16x on throughput at 75% lower cost. The practical scale ceiling is 10–50M vectors on a single instance, extendable to 100M+ with tuning or Citus sharding. The killer advantage is full SQL power: JOINs, CTEs, transactions, and the entire PostgreSQL ecosystem. Available pre-installed on Supabase ($0–25/month), Neon ($0–19/month), AWS RDS/Aurora, AlloyDB, and Timescale Cloud. ~20,400 GitHub stars. Not recommended beyond 100M vectors with low-latency requirements.
Pros
- + No new infrastructure — just a Postgres extension
- + Full SQL alongside vector search (JOINs, CTEs, ACID)
- + Available on every major managed Postgres platform
- + HNSW + DiskANN indexing (via pgvectorscale)
- + Huge PostgreSQL ecosystem
- + Dramatically improved in 2025 (v0.8.0)
Cons
- - Performance ceiling at 50–100M vectors
- - HNSW indexes must fit in RAM or latency spikes 10x
- - No built-in vectorization
- - Fewer vector-specific features than dedicated solutions
Details
- Pricing Model
- open-source
- Starting Price
- $0
- Self-Hosted
- Yes
- Cloud Hosted
- Yes
- Founded
- 2021
- Repository
- GitHub →
Best For
- • Adding vectors to existing PostgreSQL
- • Unified SQL + vector queries with JOINs and CTEs
- • Avoiding new infrastructure
- • Supabase and Neon users
Integrations
Articles featuring pgvector
Last updated: