datastax/jvector

What JVector Is

JVector is an open‑source Java library that implements a graph‑based approximate nearest‑neighbor (ANN) index. It combines ideas from two well‑known ANN algorithms – HNSW (Hierarchical Navigable Small World) and DiskANN’s Vamana – to give a fast, scalable search structure that can live partly in memory and partly on disk.

Why It Matters

Finding the k most similar vectors in a huge, high‑dimensional dataset is a core operation for many AI systems (semantic search, recommendation, clustering, etc.). Exact search is too slow at scale, so most production systems use ANN. JVector’s hybrid graph design lets you:

• Search in logarithmic time while keeping a small memory footprint.
• Build indexes larger than RAM by using a two‑pass approach – a quick, compressed first pass in memory, followed by a precise second pass that reads data from disk.
• Scale construction linearly with CPU cores thanks to non‑blocking concurrency, which is valuable when you need to index billions of vectors.

Core Features (as described in the README)

Feature	What It Does
Multi‑layer graph	Upper layers are pure in‑memory adjacency lists for ultra‑fast navigation; the bottom layer lives on‑disk.
Two‑pass search	First pass uses compressed representations (Product Quantization, Binary Quantization, or fused PQ) for speed and low memory; second pass refines results with full‑precision float32 or NVQ vectors.
Incremental construction	Index can be updated and built incrementally, unlike many partition‑based ANN methods that require a static dataset.
Non‑blocking concurrency	Construction scales linearly with core count, as shown in the README’s scaling chart.
Cross‑platform native SIMD	A native library (libjvector.so) built with Meson/Ninja provides SIMD‑accelerated distance calculations on Linux x86‑64 and AArch64.
Multi‑release JAR	Packs Java‑11 compatible code and Java‑20‑specific optimizations into a single artifact, usable from any Java 11+ project.
Benchmarks & examples	Ready‑to‑run Bench and SiftSmall classes, plus tutorial code in jvector-examples.

Who Might Use It

• Engineers building semantic search or recommendation services that need sub‑millisecond latency on billions of embeddings.
• Researchers prototyping new ANN algorithms who want a solid, production‑grade baseline written in Java.
• Companies that store embeddings on‑disk (e.g., large‑scale document or image vectors) and cannot fit the whole index in RAM.

Getting Started (quick steps)

1. Clone with submodules – the repo pulls in Google’s Highway SIMD library.
2. Build the native library – run build_native_lib.sh (Linux only) to get libjvector.so.
3. Compile the Java modules – mvn clean install (or mvn -Pjdk11 for Java 11).
4. Run a tutorial – follow docs/tutorials/1-intro-tutorial.md or inspect VectorIntro.java.
5. Benchmark – mvn compile exec:exec@bench to see performance on standard datasets.

How It Fits Into the AI Landscape

JVector isn’t a model‑training framework; it’s a search engine component that sits behind applications that already have vector embeddings (from LLMs, CLIP, sentence‑transformers, etc.). By providing a high‑performance, disk‑backed ANN index, it lets those applications retrieve nearest‑neighbors quickly without needing massive RAM, making large‑scale semantic AI systems more practical.

---

All information above is taken directly from the repository’s README.