# PixelRAG

A retrieval system that renders web pages, PDFs and images as screenshots and searches over the images directly, keeping visual structure such as tables and charts that text extraction would lose.

PixelRAG renders documents, including web pages, PDFs and images, as screenshots and retrieves over the images directly rather than over extracted text. Visual structure that HTML parsing would discard, such as tables, charts, layout and infographics, stays intact, so a downstream reader model can answer questions about it.

A hosted, pre-built index of 8.28 million Wikipedia pages is searchable through a live API with no setup or API key required, including image queries for visual search; the underlying pipeline is general-purpose and can render and index other pages or documents. A renderer command turns a page or document into screenshot tiles for a visual index.

The renderer also ships as a Claude Code plugin, so instead of fetching raw HTML, Claude Code can screenshot a page and read the resulting image, seeing charts, diagrams, tables and layout directly. It is written in Python and released under the Apache-2.0 license.

Website: https://arxiv.org/pdf/2606.28344