Tool not found.
Back to Home
Open-source visual retrieval system that renders web pages, PDFs, and images as screenshots and searches over them directly, preserving tables, charts, and layout that text parsing loses.
PixelRAG is a research codebase from Berkeley SkyLab, BAIR, and Berkeley NLP that reimagines retrieval-augmented generation around images instead of parsed text. Rather than converting documents to text chunks, which discards tables, charts, and layout, PixelRAG renders pages into screenshot tiles and embeds them with a LoRA-fine-tuned Qwen3-VL-Embedding model, so a reader model can retrieve and answer questions using the same visual structure a human would see. The project ships a standalone `pixelshot` CLI for rendering any URL or PDF into tiles, plus a modular `pixelrag` pipeline covering chunking, embedding, FAISS index building, and a FastAPI search server, each installable as separate extras. A hosted API serves a pre-built index of 8.28 million Wikipedia pages with no setup required, and the renderer also ships as a Claude Code plugin (the pixelbrowse skill) that lets Claude screenshot and visually read web pages instead of parsing raw HTML. Training code, LoRA adapters, and the training dataset are all published separately for those who want to adapt other embedding backbones.
