Postgres extension complements pgvector for performance and scale
4 months ago
- #VectorSearch
- #DatabaseExtensions
- #PostgreSQL
- pgvectorscale introduces StreamingDiskANN index, Statistical Binary Quantization, and label-based filtered vector search.
- Performance benchmarks show 28x lower p95 latency and 16x higher query throughput compared to Pinecone.
- Developed in Rust using PGRX framework, offering new contribution avenues for PostgreSQL vector support.
- Installation options include Docker, source installation, and Timescale Cloud service.
- Supports cosine, L2, and inner product distance queries with syntax similar to pgvector.
- Label-based filtering allows efficient vector searches combined with metadata filtering.
- Index build-time and query-time parameters allow customization of performance and accuracy.
- Parallel index building is supported for large datasets with configurable parameters.
- Current limitations include no support for UNLOGGED tables and relaxed ordering in results.
- Community contributions are encouraged to shape the future development of pgvectorscale.