rupixel
A Rust port of PixelRAG that searches documents and video by how a page or frame visually looks rather than only by its text, using screenshots of pages turned into visual embeddings and an approximate-nearest-neighbor index.
rupixel supports search by classic text matching as well as by taking a screenshot of each document page and searching the resulting images directly, which the project states matters when text is locked inside a scan, an image, a table or a chart that ordinary text search cannot read. It is built on the ruvector approximate-nearest-neighbor substrate, combining an HNSW index with IVF-Flat, and ported from an existing project called PixelRAG.
A real-time video-search demo samples a live feed a few times per second, skips frames that changed little through a keyframe gate, and embeds the rest with CLIP so the stream can be searched by meaning; it runs on the GPU through WebGPU when available and falls back to CPU or WebAssembly automatically. An optional streaming vision-language model, defaulting to Qwen3-VL, can narrate each keyframe as live captions with a kept timestamped transcript, describing only what changed rather than restarting the description each time.
All of the published demos run entirely in the browser with no install or server, and a bring-your-own-key mode keeps an OpenRouter key only in the browser tab's session storage; a server-side proxy is also included for a shared, managed key that never reaches the browser. It is written in Rust and released under the MIT license.