jvector

jvector

Category: SDKs & Libraries
Type: Java library for vector search and approximate nearest neighbor (ANN) indexing

jvector is a high-performance, Java-based engine for approximate nearest neighbor search, combining ideas from HNSW and DiskANN/Vamana to provide scalable, graph-based vector indexing.

Features

• Graph-based ANN index

  • Implements a graph index merging the DiskANN and HNSW design families.
  • Uses a hierarchical (multi-layer) graph structure similar to HNSW.
  • Employs the Vamana algorithm (from DiskANN) within each layer.

• Hierarchical / multi-layer architecture

  • Upper layers of the hierarchy are maintained in memory as per-node adjacency lists.
  • Enables fast navigation of the index with no disk I/O for upper-level traversals.
  • Bottom layer is stored on disk as per-node adjacency lists for large-scale datasets.

• Two-pass search support

  • Stores additional data inline in the on-disk layer to enable two-pass search strategies.
  • First pass is designed to be powered by lightweight, I/O-efficient access patterns.

• High performance and scalability

  • Non-blocking concurrency control for index construction.
  • Construction scales linearly with the number of CPU cores (parallel index build).
  • Designed for high-dimensional vector search where exact KNN is impractical.

• Approximate nearest neighbor (ANN) capabilities

  • Optimized for logarithmic-time approximate nearest neighbor search on large datasets.
  • Suitable for dynamic, general-purpose indexing where incremental construction and updates are needed.

Typical Use Cases

• Large-scale vector search over high-dimensional embeddings.
• Applications requiring fast approximate nearest neighbor queries (e.g., semantic search, recommendation, similarity search).
• Systems needing incremental or continuously updated ANN indexes rather than static, pre-built indexes.

Pricing

• Not specified in the provided content (open-source GitHub project; see repository for license and usage terms).