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 about Postgres-native ANN + BM25 inside Lakebase. It is separate from S3 Vectors metadata pre-filtering and from Aurora and DuckDB lake foreign tables.
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).
Vendor claims
All figures below are Databricks-reported—not independently 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”; we did not independently reconcile these).
- 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: serve 100M on 1 CU; storage-backed indexes survive scale-to-zero. Index builds described as offloadable / LTAP→Spark—“Stay tuned”; Spark offload is not shipped GA.
No dollar pricing is given.
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.

The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment