Neon has launched Lakebase Search, two new Postgres extensions — lakebase_vector and lakebase_text — that bring scalable ANN vector search and BM25 full-text search to Neon's storage-compute separated architecture. The extensions address key limitations of pgvector (HNSW RAM requirements, 2000-dimension cap) and PostgreSQL's native GIN-based full-text search (no true BM25 scoring, no top-K pushdown). lakebase_vector uses IVF partitioning plus RaBitQ quantization to compress vectors ~32x, enabling billion-scale indexes under 10GB. lakebase_text implements Block-Max WAND for efficient top-K retrieval with real BM25 scores. Both indexes live on object storage, making them compatible with Neon's scale-to-zero and instant branching features. Hybrid search is expressed as plain SQL within a single Postgres transaction, eliminating the need for separate vector databases or search clusters.
