Repolode
Submit
StarTrail-org

StarTrail-org/PixelRAG

Open-source pipeline that renders web pages and PDFs as screenshots and retrieves them visually, so tables and charts survive for RAG.

View on GitHubWebsite
Stars
10,309
Forks
910
Open issues
19
Last commit
5 days ago

What is it?

PixelRAG treats documents as images rather than text. It renders web pages, PDFs and images into screenshot tiles, embeds them with a Qwen3-VL-Embedding model, and searches a FAISS or Qdrant index. Because layout, tables, charts and infographics are kept intact, a reader model can answer questions that text-chunk pipelines lose when HTML is parsed.

The project ships several pieces: the pixelshot CLI for capture via Playwright/CDP, the pixelrag umbrella command for chunking, embedding, index building and a FastAPI search server, and a LoRA fine-tuning recipe in a separate train/ project. A hosted search API over a pre-built index of 8.28M Wikipedia pages is available, and the same rendering can be used by Claude Code or opencode through a plugin.

Key features

  • 01Renders web pages and PDFs into screenshot tiles with pixelshot.
  • 02Embeds page images with a Qwen3-VL-Embedding model that can be LoRA fine-tuned.
  • 03Builds FAISS or Qdrant indexes and serves search through a FastAPI endpoint.
  • 04Supports image queries for visual search.
  • 05Runs on Linux with CUDA and on macOS with Apple Silicon MPS.

Use cases

  • Answering questions over documents where tables and charts carry the key numbers.
  • Letting an AI coding agent read a web page as an image instead of raw HTML.
  • Indexing a private collection of PDFs for visual search on a local machine.

Topics

#rag#multimodal#vision#vector-search#embeddings#screenshot#faiss#qdrant#agent#ai

FAQ

What is PixelRAG?
Open-source pipeline that renders web pages and PDFs as screenshots and retrieves them visually, so tables and charts survive for RAG. PixelRAG treats documents as images rather than text. It renders web pages, PDFs and images into screenshot tiles, embeds them with a Qwen3-VL-Embedding model, and searches a FAISS or Qdrant index. Because layout, tables, charts and infographics are kept intact, a reader model can answer questions that text-chunk pipelines lose when HTML is parsed.
Is PixelRAG free to use? What license does it have?
PixelRAG is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is PixelRAG still maintained?
Actively maintained. The last commit to PixelRAG was on Oct 14, 2026.
What can I use PixelRAG for?
Answering questions over documents where tables and charts carry the key numbers. Letting an AI coding agent read a web page as an image instead of raw HTML. Indexing a private collection of PDFs for visual search on a local machine.
Where can I get PixelRAG?
The source code and installation instructions are on GitHub: https://github.com/StarTrail-org/PixelRAG

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects