vlite is a small, fast vector database implemented in pure NumPy — no server to run and no database to set up. Published on GitHub with about 770 stars under the MIT license and updated in 2026, it is designed to make embedding search feel as portable as a browser cookie.
Vlite stores embeddings in a portable CTX (context) file that you can take anywhere, making it aimed squarely at retrieval-augmented generation and personal context use cases. Embeddings from the mixedbread embed-xsmall model are baked in, so adding text is a single call; long texts are automatically chunked to the model’s 512-token window, and metadata filtering is built in. The project reports strong benchmark results against heavier alternatives, claiming sub-1.1-second search over 500k documents with binary embeddings and a much smaller on-disk footprint than Chroma in its tests.
The usage model mirrors that simplicity. You create a VLite instance with a name, add a list of texts, and immediately query it: results come back with the matched text and a distance score, ready to feed into a prompt. Because the whole store lives in a single CTX file, a knowledge base can be copied, versioned, or shipped alongside an application like any other asset — an approach the author compares to browser cookies, but for embedding-based context.
vlite is free and open source under the MIT license. There are no paid tiers, hosted plans, or usage limits; it runs entirely on your own machine.
vlite fits developers prototyping RAG features, building local-first AI tools, or embedding semantic search into scripts and small apps who find full vector database servers unnecessarily heavy.