Timescale Releases pg_vectorsearch: A PostgreSQL ANN Index Faster Than pgvector's IVFFlat and HNSW

Timescale has released pg_vectorsearch, an alpha-stage PostgreSQL extension for high-performance approximate nearest neighbor (ANN) search. The tool introduces the PRISM index, which combines IVF-style clustering with hierarchical routing to support datasets exceeding 100 million vectors. It leverages RaBitQ quantization for two-stage search with theoretical error bounds and utilizes SIMD-optimized distance calculations (AVX2, AVX512, NEON).

Performance and Architecture

The project claims significantly faster query and index build times compared to pgvector’s IVFFlat and HNSW implementations. In benchmarks on a 100M vector dataset, pg_vectorsearch reportedly outperforms graph-based HNSW at equivalent recall levels while using only 1/4 of the memory. The architecture relies on a hierarchical centroid tree for clustering and stores RaBitQ-quantized vectors in posting lists optimized for SIMD operations.

Compatibility and Usage

pg_vectorsearch is binary-compatible with existing pgvector columns, allowing users to create PRISM indexes on vector and halfvec types without rewriting data. It supports standard operators like <-> (L2), <#> (Inner product), and <=> (Cosine). Key tuning is managed via the prism.nprobe parameter, where a value of 0 enables auto-tuning for ~0.95 recall.

Requirements and Status

The extension requires PostgreSQL 18+, a GNU C11 compiler (GCC 11+ or Clang 14+), and the Meson build system. As an alpha release, it is not suitable for production; upgrades between prereleases may require dropping and recreating indexes. Dynamic updates are supported, but posting list splits must be manually triggered via the prism_rebalance procedure.