Lakebase Search Launches with Enhanced Capabilities

Lakebase Search is now generally available on AWS and Azure, delivering twice the throughput of the next-best system on a 100M-vector benchmark. This performance is particularly notable as Lakebase achieves a 97% recall rate with a P99 latency of just 71ms on the LAION 100M dataset.

The system includes two key extensions: lakebase_vector, which provides approximate nearest neighbor (ANN) search, and lakebase_text, which implements the BM25 algorithm for text search. Conexiom, a customer utilizing Lakebase, reported a threefold reduction in infrastructure costs and five times higher throughput compared to using pgvector.

Jordan Voves, AI/ML Architect at Conexiom, emphasized the scalability benefits, stating that Lakebase Search unlocks BM25 capabilities within a serverless database architecture. This allows for seamless scaling from 1 row to 1 billion vectors, addressing both operational and search workloads effectively.

In terms of resource efficiency, Lakebase requires only 3.3 KB of memory for a 768-dimensional float32 vector and necessitates 330 GB of RAM to keep the index resident for millisecond query responses at 100 million rows. Notably, it takes 50 hours to build a pgvector index on a standard cloud instance, highlighting Lakebase's efficiency in comparison.

The system's performance was benchmarked against DiskANN and evaluated using VectorDBBench, confirming its competitive edge. With Lakebase capable of serving 100 million vectors on a single compute unit, it positions itself as a robust solution for organizations looking to enhance their search capabilities while reducing costs significantly.