Show HN: NanoVector – A 120KB zero-dependency vector search engine in C and SIMD (github.com)

NanoVector, a new zero-dependency vector search engine written in bare-metal C, has been announced, offering a streamlined alternative to cumbersome vector databases like ChromaDB and FAISS. With a size of just about 120KB and leveraging advanced SIMD instructions (AVX2, FMA, ARM NEON), NanoVector provides sub-millisecond search times and significantly improved efficiency. It addresses common issues in AI agents and local LLM pipelines, such as cold start penalties and high memory overhead, by using brute-force SIMD search techniques that keep data operations within CPU cache. This tool is particularly significant for the AI/ML community, as it enables developers to handle up to 1.4 million vector ingestions per second and achieve search latencies of just 0.13 ms for querying 2,000 vectors. NanoVector’s unique architecture allows for instant installation, usage with Python, and efficient memory management without additional dependencies. Its single-file storage format aids in fast persistence and reloading of vector data. Overall, NanoVector stands to enhance the performance and usability of AI applications, especially those involving episodic memory and dynamic vector storage, while providing a lean alternative to traditional solutions.