unum-cloud/USearch
USearch
USearch is a high-performance vector similarity search engine and approximate nearest neighbor index designed for dense embeddings. It functions as a low-level vector database core and high-dimensional vector indexer, providing the primitives necessary to store and retrieve vectors across massive datasets.
The engine distinguishes itself through hardware-level SIMD acceleration for distance kernels and a proximity-graph indexing system that enables fast retrieval across billions of vectors. It supports multi-precision vector quantization to balance memory usage and accuracy, and utilizes memory-mapped index persistence to reduce RAM overhead during loading and serialization.
The project covers a broad range of capabilities including exact brute-force linear scans, batch processing for bulk similarity searches, and thread-safe concurrent index construction. It implements multiple distance metrics—such as Euclidean, Hamming, Jaccard, and Haversine for geospatial proximity—while allowing for the integration of custom user-defined metric functions.
Additional utility surfaces include vector data clustering, semantic data joining, and tools for benchmarking search performance and accuracy evaluation.
Features
- Approximate Nearest Neighbor Search - Ships a proximity-graph based indexing system for fast approximate nearest neighbor searches across massive datasets.
- Graph-Based Indexing - Organizes high-dimensional vectors into a graph structure to enable fast approximate nearest neighbor search via traversal.
- Vector Similarity Search - Implements a high-performance engine for finding nearest neighbors in high-dimensional datasets using SIMD-accelerated distance metrics.
- Proximity Graph Indexes - Uses a proximity-graph indexing system to enable fast retrieval across billions of vectors.
- Index Construction - Provides thread-safe concurrent insertion and pre-allocated memory management for constructing vector indices.
- Memory-Mapped Indexing - Maps binary index files directly into the virtual address space to enable fast loading and low RAM overhead.
- Serializable Indexes - Writes the proximity graph and metadata to a binary file and reads it back without requiring a full memory load.
- Similarity Search - Implements both approximate graph-based and exact SIMD-optimized similarity search methods.
- Similarity Search Engines - Provides a computational engine optimized for identifying nearest neighbors in high-dimensional spaces using advanced indexing.
- Vector Databases - Functions as a low-level storage and retrieval system for dense embeddings with quantization and memory-mapping.
- Vector Indexing - Provides tools for creating and managing high-dimensional vector indexes with support for binary serialization and memory-mapped storage.
- Vector Quantization - Supports storing vectors in various bit-depths from double precision to 1-bit binary to balance accuracy and memory usage.
- High-Dimensional Vector Compressors - Reduces the memory footprint of vector collections through quantization and precision casting.
- SIMD-Accelerated Arithmetic - Computes vector similarities using hardware-level SIMD instructions for high-throughput numerical processing.
- Vector Distance Kernels - Ships hardware-level SIMD acceleration for Euclidean, cosine, and Hamming distance kernels.
- Vector Distance Metrics - Provides SIMD-accelerated mathematical primitives for calculating distances between high-dimensional vectors.
- Thread-Safe Index Construction - Utilizes a thread-safe processing engine to run multiple additions and searches across different threads simultaneously.
- Billion-Scale Vector Search - Scales similarity search to billions of vectors using custom integer types and optimized memory management.
- Batch Processing - Allows adding multiple concatenated vectors in a single operation to increase indexing throughput.
- Brute-Force Search Algorithms - Performs an exact search by calculating distances across every vector in a dataset using parallel priority queues.
- Semantic Joins - Creates mappings between two large vector datasets using approximate similarity to perform fuzzy or semantic joins.
- Geospatial Search - Calculates real-world distances between spherical coordinates using Haversine metrics for location-based retrieval.
- Quantized Optimizers - Implements floating-point and integer quantization to balance search accuracy and memory consumption.
- Parallel Index Lookups - Supports simultaneous searching across multiple indices to handle workloads containing trillions of vectors.
- User-Defined Functions - Allows the integration of custom compiled functions or assembly code to implement specialized distance calculations.
- Vector Index Compression - Reduces memory footprint by storing vectors as half-precision floats, 8-bit integers, or packed binary formats.
- Batch Vector Queries - Enables simultaneous querying of multiple vectors using tensors to increase throughput for bulk workloads.
- Brute Force Linear Scans - Provides exact vector search via brute-force linear scans for guaranteed accuracy on smaller datasets.
- User-Defined Distance Metrics - Supports the integration of custom compiled functions or assembly code to implement specialized distance calculations.
- Vectorized Data Processing - Processes multiple query vectors simultaneously using flattened arrays to maximize throughput for bulk similarity searches.
- Performance Benchmarks - Includes tools for tracking search efficiency and operations per second to analyze index performance.
- Concurrent Index Construction - Uses a multi-threaded execution engine to distribute vector additions and index construction across multiple CPU cores.
- Search Accuracy Evaluation - Measures index quality by comparing approximate results against exact ground-truth searches using recall metrics.
- AI & Machine Learning - Search and clustering engine for vectors.
- Machine Learning - Fast search and clustering for vectors.
- Database Systems - Similarity search engine for vectors and strings.
- Databases & Data - Similarity search engine for vectors and strings.
