PixelRAG

PixelRAG is a visual retrieval-augmented generation pipeline that renders documents as screenshot tiles and retrieves over those images, rather than parsing text. Standard RAG pipelines parse HTML or PDF to plain text, discarding tables, charts, and visual layout in the process. PixelRAG keeps that structure intact by treating every page as an image, then using a vision-language embedding model to match queries against the visual content. It was developed at Berkeley SkyLab, BAIR, and Berkeley NLP.

Install the base package with `pip install pixelrag` (requires Python 3.12+), which gives you the `pixelshot` CLI for rendering pages to tiles. For a full local pipeline, install the index extra with `pip install 'pixelrag[index]'`, write a `pixelrag.yaml` config pointing at your documents, run `pixelrag index build`, then serve with `pixelrag serve`. You can also query the hosted Wikipedia index immediately, no setup or API key needed, via `curl -X POST https://api.pixelrag.ai/search`.

Yes, PixelRAG is released under the Apache-2.0 license and the full codebase is on GitHub. The hosted `api.pixelrag.ai` endpoint serving 8.28M Wikipedia pages is also free and requires no API key. Self-hosting your own index requires your own hardware and roughly 217 GB of disk for the base Wikipedia FAISS index.

PixelRAG is strongest when your documents are visually dense: pages with tables, charts, infographics, or layouts that text parsers would mangle. It is also a good fit for giving Claude Code visual browsing capability via the `pixelbrowse` plugin, letting Claude see charts and diagrams the way a person does instead of reading raw HTML. The hosted Wikipedia API makes it easy to validate the approach on a real-world scale before committing to a self-hosted index.

Text-based RAG converts documents to plain text chunks, which silently drops tables, charts, and layout; PixelRAG renders to screenshot tiles and retrieves the images directly, preserving that visual structure. The tradeoff is infrastructure: PixelRAG requires a vision-language embedding model (Qwen3-VL-Embedding-2B), a FAISS index, and a rendering step, whereas text RAG pipelines are lighter to self-host. Per the README's research title, the Berkeley team found that screenshot-based retrieval beats text retrieval for web documents.

PixelRAG requires Python 3.12+ and works best with a GPU or Apple Silicon for embedding at reasonable speed, the README notes roughly 3 minutes per PDF on Apple M-series and 1 minute on GPU. The bundled turbo headless Chrome auto-installs only on linux-x64; macOS and Windows users need a system Chrome or Playwright's Chromium already installed. Downloading the base Wikipedia FAISS index locally is approximately 217 GB, making full self-hosting storage-intensive. The training environment pins specific versions (torch 2.9.1, transformers 4.57.1) in a separate project, though you do not need to retrain to use the published adapters.