Elastic launches a serverless vector database

Elastic launched a serverless vector database that supports workloads from RAG and agent memory to hundreds of billions of vectors, while its auto-calibration tests averaged 16.7% higher QPS across 18 datasets. Elastic’s announcement says vectordb_document defaults to bfloat16 and Better Binary Quantization can shrink vector memory footprints by up to 32x; those are vendor results, not an independent benchmark. Builders with retrieval-heavy agents should price migration against the archive’s skill-retrieval ceiling: managed defaults remove plumbing, but relevance, tenant isolation, and recall still need workload tests.