PixelRAG

PixelRAG is the code release for the paper "PIXELRAG: Web Screenshots Beat Text for Retrieval-Augmented Generation" from Berkeley's Sky Computing Lab, BAIR, and Berkeley NLP. It renders documents (web pages, PDFs, and images) as screenshot tiles and retrieves over the images directly, so tables, charts, layout, and infographics that HTML-to-text parsing discards remain available to the reader model.

Two pieces make this work: rendering documents to images instead of parsing them, and a Qwen3-VL-Embedding model fine-tuned with LoRA on screenshot data so that visual content is retrievable. The pipeline runs in stages: capture pages into tiles with headless Chromium, chunk and embed the tiles, build a FAISS index, and serve a search API on CPU or GPU. Indexing runs on Linux with CUDA or on Apple Silicon.

A hosted endpoint serves a pre-built index of 8.28 million Wikipedia pages with no API key and accepts text or image queries, and the pre-built FAISS indexes (Wikipedia pixel and text, and a news pixel index) are published on Hugging Face. The training code, data, and synthetic query pipeline are released so other embedding backbones can be adapted.

The renderer also ships as the pixelbrowse skill for Claude Code and as a plugin for OpenCode. Instead of fetching raw HTML, the agent screenshots a page and reads the image, with no MCP server or backend involved.