Timescale's alpha PostgreSQL extension claims lower-memory vector search

Timescale's pg_vectorsearch extension offers an alpha PostgreSQL index whose maintainers say outperforms pgvector's HNSW index at equivalent recall on a 100-million-vector dataset using an instance with a quarter of the memory. Its PRISM index searches for similar vectors by routing queries through clusters and filtering compressed candidates before checking full-precision values. The extension can index existing pgvector columns without rewriting tables or changing queries, and requires PostgreSQL 18 or later. It supports inserts, updates and deletes. The maintainers designate it for testing rather than production and warn that prerelease upgrades may require dropping and recreating the extension and its indexes.