hora

As embedding-based retrieval has become central to semantic search, recommendation systems, and retrieval-augmented generation, the need for ANN libraries that can scale to millions of vectors while maintaining low query latency has grown significantly.

Hora implements multiple indexing algorithmsHNSW, IVFPQ, and othersbehind a unified API, enabling developers to choose the right trade-off between speed, memory, and recall for their specific use case.

The Rust implementation gives Hora strong performance characteristics and memory safety without garbage collection overhead, making it suitable for latency-sensitive production services where JVM pauses or Python GIL contention would be problematic.

Hora exposes Python bindings alongside its Rust API, allowing data scientists to prototype with a familiar interface while the underlying computation runs in native Rust.

The library also supports WebAssembly compilation, enabling vector search to run directly in browser environmentsan unusual capability that opens use cases in client-side semantic search without server round-trips.

Developers building semantic search features, recommendation engines, or RAG pipelines use Hora when they need an embeddable vector index that avoids the operational overhead of dedicated vector database services.

For applications where the vector index can be loaded into process memory alongside application logic, Hora provides query performance competitive with dedicated services while eliminating the network latency of external database calls.