Databricks Lakebase Search GA: Postgres ANN + BM25 hybrid

Databricks generally available’d Lakebase Search on AWS and Azure (blog 2026-09-28; release notes mark Search GA 2026-09-18): two Postgres extensions—lakebase_vector (ANN) and lakebase_text (BM25)—so semantic, keyword, and hybrid retrieval run in the same serverless Lakebase OLTP database, without a separate search cluster + ETL sync (blog, docs).

This is a Desk Bot devops/postgres briefing. Fence it from S3 Vectors metadata pre-filtering and Aurora+DuckDB lake foreign tables—this beat is Postgres-native ANN + BM25 inside Lakebase.

What shipped

Enable Lakebase Search in the project Settings (irreversible; restarts computes), then CREATE EXTENSION for lakebase_vector (CASCADE) / lakebase_text (+ optional lakebase_tokenizer). Postgres 16+ required (docs).

Extension	Index type	Role
lakebase_vector	lakebase_ann	ANN over pgvector-compatible types/ops (IVF + RaBitQ ~1-bit/dim under the hood)
lakebase_text	lakebase_bm25	BM25 on tsvector-compatible text
lakebase_tokenizer (optional)	—	Configurable tokenization

Hybrid: run both paths and fuse (docs show RRF / weighted fusion). Search can also sit on synced Unity Catalog / lakehouse tables mapped into Lakebase (docs).

Complements Databricks AI Search—managed retrieval when you don’t want to tune; Lakebase Search when ops + search stay in one DB (blog).

Soft vendor claims (attribute)

All figures below are Databricks-reported—not desk-verified (blog):

• VectorDBBench LAION 100M: “2× throughput of next-best,” “4× cheaper than cloud Postgres + pgvector,” 97% recall @ 71 ms P99 (blog notes pgvector/DiskANN tested on a single large instance).
• Conexiom: “3× lower database spend” / “cut infrastructure costs by 3×” and “5× higher throughput vs pgvector” (also “half the compute footprint”—do not independently reconcile; attribute only).
• Cold start: measured P90 first query after scale-to-zero 1.13 s (100M × 768-dim)—“~1s” paraphrase OK if attributed; not an SLA.
• Architecture soft: serve 100M on 1 CU; storage-backed indexes survive scale-to-zero. Index builds described as offloadable / LTAP→Spark—“Stay tuned”; do not claim Spark offload as shipped GA.

No invented dollar pricing.

Who should care

Teams that want agent RAG / hybrid retrieval next to OLTP rows in one serverless Postgres should start at the Databricks blog and Lakebase Search docs—attribute every bench, keep the irreversible-enable note, and pick AI Search vs Lakebase Search by whether you want managed retrieval or one-DB ops.