Pg_vectorsearch: An index that is faster than pgvector's IVFFlat and HNSW(github.com)
206 points by jascha_eng 1 day ago | 52 comments
tl;dr: Pg_vectorsearch is an alpha-stage PostgreSQL extension providing PRISM, an IVF-style ANN index that uses RaBitQ quantization, hierarchical centroid routing, and SIMD-optimized distance computation to outperform pgvector's IVFFlat and HNSW on both query and build times (reportedly faster than HNSW at equivalent recall on 100M vectors with a quarter of the memory). It's compatible with existing pgvector `vector` and `halfvec` columns, so it can be installed alongside pgvector without rewriting data or queries. Requires PostgreSQL 18+ and offers native `vec32`/`vec16` types for new deployments.
HN Discussion:
  • ~Practical adoption limitations like 1968 dimension cap and PG18+ requirement constrain real-world usage
  • ↑Excitement about competition in low-RAM vector indexes leveraging fast NVMe drives
  • ↓Concerns about missing filtering support for multi-tenant/permission scoping
  • ↓Skepticism about the team abandoning prior projects like pgvectorscale rather than improving them
  • ~Adoption depends on managed cloud Postgres providers supporting the extension