Lakebase Search: vector and BM25 on Neon

Lakebase Search, now available on Neon, offers a hybrid vector and full-text retrieval system through two Postgres extensions: lakebase_vector and lakebase_text. These extensions are tailored for Neon's architecture, which separates compute from storage, and they optimize vector similarity and BM25 keyword searches while maintaining compatibility with existing Postgres queries. Lakebase_vector supports scalable vector similarity search using inverted file partitioning and RaBitQ quantization, enabling it to handle over a billion vectors efficiently. Meanwhile, lakebase_text enhances BM25 keyword search with native indexing that includes corpus-wide statistics and top-K pushdown for efficient query performance. By leveraging Neon's tiered storage system, these extensions provide durable and scalable search capabilities without the need for additional database systems, allowing hybrid retrieval directly within Postgres. The design ensures that indexes remain durable on object storage, facilitating scale-to-zero operations and enabling branches for testing and tuning search configurations without affecting production environments.

No tracked trend matches for this post yet.