datastax/jvector

Implements a hybrid graph-based index combining HNSW's hierarchical structure with Vamana's algorithms, using a two-pass search strategy with in-memory compressed vectors (PQ/binary quantization) for fast filtering and disk-resident full-precision vectors for accuracy. Supports incremental construction with nonblocking concurrency that scales linearly across cores, and uniquely enables building larger-than-memory indexes by applying two-pass search during index construction itself. Built as a Java library with multirelease JAR support for Java 11+ and optional SIMD optimizations on Java 20+.

1,693 stars. Actively maintained with 5 commits in the last 30 days.