Powering AI at Scale: Benchmarking 1B Vectors in YugabyteDB (www.yugabyte.com)

YugabyteDB announced a large-scale vector search benchmark: indexing and querying 1 billion 96-dimensional embeddings from the Deep1B dataset, achieving 96.56% recall with median query latency of 0.319 seconds. The system used HNSW with m=32, ef_construction=256 and ef_search=256, showing that a distributed SQL database can deliver high-accuracy, sub-second approximate nearest neighbor (ANN) search at billion-vector scale. Technically, this was achieved by combining Usearch (an optimized vector engine) with a Vector LSM abstraction that stages vectors in in-memory buffers and immutable on-disk chunks, plus automatic sharding and shard redistribution to scale memory and compute horizontally. Vectors are co-located with relational rows, queries run in parallel across tablets with local top-K aggregation, and multiple smaller indexes are managed automatically to keep latency low. Yugabyte’s pluggable index architecture and PostgreSQL compatibility (pg_vector extension and reused Postgres query engine) let teams run transactional and vector workloads in one system, while tuning HNSW parameters and node count balances recall, latency, and memory use. The result demonstrates a practical path for enterprises to unify RAG, recommendations, and semantic search without deploying separate vector stores.