Google Cloud previewed native BM25 keyword ranking in AlloyDB and Cloud SQL for PostgreSQL 17+, eliminating separate full-text search backends.

Google announced preview support for native Best Matching 25 (BM25) indexing in AlloyDB and Cloud SQL for PostgreSQL 17+. The implementation uses Tiger Data’s open-source pg_textsearch extension, bringing C-optimized keyword scoring directly to PostgreSQL tables.

The stated application is hybrid search: combining semantic vector retrieval with exact keyword full-text retrieval. Google notes that vector embeddings can struggle with specific alphanumeric identifiers and exact product stock-keeping-unit numbers, even when they capture conceptual meaning. Native BM25 makes those retrieval methods available within the operational database.

Google contrasts BM25 with PostgreSQL’s built-in `ts_rank`, highlighting inverse document frequency, term-frequency saturation and document-length normalization. These mechanisms weight rare terms, prevent repeated words from dominating scores and account for document length. Google says the integrated approach removes the need for a separate full-text backend and its associated duplication, synchronization lag and extract-transform-load pipelines.

Separately, the announcement cites AlloyDB-exclusive vector-query performance of up to 6× and 10× faster than standard PostgreSQL, using ScaNN and hierarchical navigable small world (HNSW) index types, respectively. Those are vector-search claims, not BM25 benchmark results. The new BM25 capability remains in preview.