USearch

USearch is an open-source similarity search and clustering engine from Unum Cloud. It implements the HNSW algorithm in a single C++11 header library and ships native bindings for languages including Python, JavaScript, Rust, Java, Go, C#, Swift and others. The project is released under the Apache 2.0 license.

What It Is

USearch is a vector search engine for approximate nearest-neighbor lookup. You create an index with a chosen dimensionality, metric and storage type, add vectors under integer keys, and query for the nearest matches. It also supports exact brute-force search for smaller collections, K-Nearest Neighbors clustering with sub-clustering, and approximate fuzzy or semantic joins between two indexes.

Metrics, Precision and Memory

Beyond the standard cosine, inner-product and L2 metrics, USearch accepts user-defined metrics, including ones written in Python with Numba and JIT-compiled. Haversine distance and binary Tanimoto and Sorensen coefficients are available for geospatial, chemistry and genomics work. Vectors can be stored in types from f64 down to bf16, f16, Float8, Float6, i8 and single-bit, and a uint40 neighbor type allows indexes beyond 4 billion entries. Predicate functions can filter results during graph traversal.

Serving and Deployment

Indexes can be saved to files, streams or buffers, and large indexes can be viewed from disk through memory mapping without loading them into RAM. Multiple smaller indexes can be searched together for very large workloads. An index built in one language can be queried from another. The README lists Linux, macOS, Windows, iOS, Android, WebAssembly and SQLite as supported targets.

Performance Claims and Integrations

Unum states that USearch is faster than FAISS, reporting for example 105,000 vec/s versus 5,500 vec/s on 96-dimensional vectors in its own benchmarks. The README lists integrations or usage in ClickHouse, DuckDB, ScyllaDB, TiDB, YugabyteDB, Memgraph, LangChain, Semantic Kernel and Sentence-Transformers. Documented examples include multimodal text-to-image search with UForm and UCall, molecular search with RDKit, and geospatial search.