NMSLIB

USearch is a lightweight, header-only C++ library for ANN search with HNSW and scalar quantization, optimized for low RAM and high-speed on CPU. It supports binary and custom metrics for edge devices. Compared to Faiss, USearch is simpler, faster on small datasets, and embeddable without deps.

Annoy (Approximate Nearest Neighbors Oh Yeah) is a pure ANN index library implementing random projection trees for fast approximate nearest neighbor search in read-heavy workloads with static indexes. Features C++/Python bindings, multi-threading, memory mapping; no quantization or GPU support. Ideal for custom vector engines, benchmarks, and low-latency recommendations; lightweight building block vs full vector DBs like Qdrant.

DiskANN is a pure ANN index library implementing Vamana graphs for disk-based billion-scale approximate nearest neighbor search with low memory footprint. Features GPU acceleration, dynamic updates, cached SSD search, C++/Python bindings. Suited for custom vector engines handling large cold datasets in search/recommendations, benchmarks; more disk-efficient than HNSWLib vs full DBs like Qdrant.

HNSWLIB is a pure ANN index library implementing Hierarchical Navigable Small World (HNSW) graphs for high-performance approximate nearest neighbor search. Features L2/cosine metrics, multi-threading, low memory, C++/Python bindings. Ideal for custom vector engines, benchmarks on millions of vectors; core building block for DBs like Qdrant/Chroma vs complete solutions.

nanoflann is a pure ANN index library implementing KD-trees for nearest neighbor search, header-only C++11 optimized for 2D/3D point clouds. Features efficient spatial queries, no quantization or GPU support, easy integration. Suited for custom vector engines in robotics and computer vision, benchmarks; lightweight building block vs full DBs like Qdrant.

ScaNN (Scalable Nearest Neighbors) is a pure ANN index library using anisotropic vector quantization and scorers for high-recall, high-throughput search at billion-scale. Features CPU/GPU support, TensorFlow/Numpy bindings, advanced quantization. For custom vector engines in recommendations, benchmarks; superior recall/throughput vs Faiss, building block unlike full Qdrant.