Get the Visual RAG Playbook

Classic RAG parses every web page down to plain text — and silently destroys the tables, charts, layouts and infographics BEFORE retrieval even starts. PixelRAG, out of Berkeley's Sky Lab, never parses: it screenshots the page, slices it into tiles, and embeds the PIXELS, so you search a document by how it LOOKS, not just the text it contains. This field guide is the whole thing, step for step: why text-RAG is structurally blind (a table becomes a flat run-on string, a chart becomes nothing), how PixelRAG works across its five stages (render with pixelshot → chunk into tiles → embed with the fine-tuned Qwen3-VL vision model → FAISS index → serve), and the three ways to use it TODAY — the free hosted index of 8.28M Wikipedia pages with one curl and no API key, the Claude Code pixelbrowse plugin that lets Claude actually see charts and tables, and indexing your own PDFs locally. Plus the public benchmarks (SimpleQA 78.8% vs 71.6%, table queries 48.8% vs 42.5%), a visual-vs-text decision chart for when to reach for each, and the honest caveats. Open source, Apache-2.0. Give your AI eyes.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.