PixelRAG
PixelRAG is a visual retrieval-augmented generation system from Berkeley SkyLab, BAIR, and Berkeley NLP. It renders web pages and documents as screenshots and retrieves over the images directly, rather than parsing HTML or extracting text. The core claim, from the accompanying paper ("Web Screenshots Beat Text for Retrieval-Augmented Generation"), is that visual structure -- tables, charts, layout, infographics -- preserved in screenshots produces better retrieval than text extracted from the same pages.
The pipeline has two operations:
-
Render:
pixelshotconverts a URL or document to screenshot tiles. - Search: A visual index is queried by text or by image. The hosted API serves a pre-built index of 8.28M Wikipedia pages with no setup or API key required.
PixelRAG also ships as a Claude Code plugin (pixelbrowse skill). When installed, Claude screenshots pages with pixelshot and reads the image rather than fetching raw HTML, giving it access to visual content that HTML parsing discards.
GitHub: StarTrail-org/PixelRAG. Created May 29, 2026. 5,009 stars, Apache-2.0, Python. Live demo at pixelrag.ai; live API at api.pixelrag.ai. CI badge passing. Academic origin: equal contribution from Yichuan Wang and Zhifei Li; advised by Matei Zaharia, Joseph Gonzalez, and Sewon Min.
PixelRAG's visual retrieval approach is adjacent to the Semantic Frame Scoring pattern: both use visual representations (screenshots vs. video frames) as the retrieval unit rather than extracted text or metadata. The pixelbrowse Claude Code plugin is directly usable in this workflow for sourcing pipeline documentation from pages with heavy visual structure.
Placed in Tools/Assess at inner because the Berkeley/BAIR academic origin, Apache-2.0 license, live hosted API, and passing CI indicate a credible and usable project. The visual-first retrieval approach is differentiated. Inner position reflects the strong production signal (live API, live demo). No first-person production use yet.
