• Sources: primary, discussion
  • Summary: PRISM is an IVF-style index with a hierarchical centroid tree kept in the buffer cache, RaBitQ-quantized posting lists, and full-precision reranking from the heap. The project claims it beats pgvector IVFFlat and HNSW on query and build speed, including at equal recall on 100 million vectors with a quarter of the memory. It requires PostgreSQL 18, is marked alpha and not for production, and may require dropping indexes between prereleases.
  • Why it matters: Teams on pgvector can test a PRISM index beside an existing HNSW index without rewriting tables or queries, but should wait for a stable release.