JVector: Cutting-Edge Vector Search in Java

An airhacks.fm conversation with Jonathan Ellis (@spyced) about:

discussion of JVector, a Java-based vector search engine, Apache Kudu as an alternative to Cassandra for wide-column databases, FoundationDB - is a NoSQL database, explanation of vectors and embeddings in machine learning, different embedding models and their dimensions, the Hamming distance, binary quantization and product quantization for vector compression, DiskANN algorithm for efficient vector search on disk, optimistic concurrency control in JVector, challenges in implementing academic papers, the Neon database, JVector's performance characteristics and typical database sizes, advantages of astra DB over Cassandra, separation of compute and storage in cloud databases, Vector's use of Panama and SIMD instructions, the potential for contributions to the JVector project, Upstash uses of JVector for their vector search service, the cutting-edge nature of JVector in the Java ecosystem, the logarithmic performance of JVector for index construction and search, typical search latencies in the 30-50 millisecond range, the young and rapidly evolving field of vector search, the self-contained nature of the JVector codebase

Jonathan Ellis on twitter: @spyced