If the benchmarks are to be believed, and assuming ~0.93 recall is "good enough" for the use case, this vector DB is seriously fast: about 5x the queries/sec of OpenSearch and 19x Milvus.