Skip to content
HUIAHUB

StarTrail-org/PixelRAG

PixelRAG

https://arxiv.org/abs/2606.28344. The end of web parsing. The beginning of scalable pixel-native search. link: https://pixelrag.ai/

Metrics updated Established
View on GitHub (opens in a new tab)Homepage / demo

GitHub metrics

Stars
10K
growth after 7 days of data
Forks
901
30-day growth
—
Open issues
24

The chart appears once the repository has been checked on two or more days.

What it does

PixelRAG is a retrieval system that turns web pages, PDFs and images into screenshot tiles and searches those images instead of extracted text. It is the official code for the arXiv paper "PIXELRAG: Web Screenshots Beat Text for Retrieval-Augmented Generation", written by researchers at Berkeley SkyLab, BAIR and Berkeley NLP.

The pipeline runs in stages. pixelshot renders a document into tiles with a headless Chromium browser driven over CDP. The pixelrag command then chunks and embeds the tiles with Qwen3-VL-Embedding and LoRA adapters fine-tuned on screenshot data, builds a FAISS index, and serves search through a FastAPI app on CPU or GPU. Each stage is its own pip extra, so a project installs only the parts it uses.

The team also runs a hosted index of 8.28M Wikipedia pages behind a live API, which takes text or image queries; the README says it needs no API key. The renderer doubles as a Claude Code plugin called pixelbrowse, which has Claude screenshot a page and read the image, and as an npm package for opencode. Training code sits in a separate train/ project with its own pinned PyTorch and transformers versions.

Why it matters

Text-based RAG parses a page into chunks and can lose tables, charts and layout along the way. PixelRAG keeps each page as an image, so a vision-capable reader model can take a number straight off a table tile. The paper's title states the core claim: web screenshots beat text for retrieval-augmented generation. A ready-made Wikipedia index means you can test the approach before building your own.

Main features

  • `pixelshot` renders web pages and PDFs into screenshot tiles
  • Embeds tiles with Qwen3-VL-Embedding plus screenshot-tuned LoRA adapters
  • Builds FAISS indexes, with an optional Qdrant backend
  • `pixelrag serve` exposes a FastAPI search API on CPU or GPU
  • Hosted Wikipedia index of 8.28M pages, searchable by text or image
  • pixelbrowse plugin for Claude Code and an npm package for opencode
  • Separate pip extras for each pipeline stage

Who it's for

  • RAG builders whose documents carry tables, charts or infographics that text parsing drops
  • Researchers reproducing or extending the PixelRAG paper
  • Claude Code and opencode users who want an agent to see rendered pages

Strengths

  • Keeps tables, charts and layout available to the reader model
  • The hosted index lets you try search without building anything
  • Modular install, with one extra per pipeline stage
  • Training code and the training dataset are published for reproduction

Limitations

  • The base Wikipedia index download is listed at about 217G
  • Outside linux-x64 you must install Chrome yourself or set `CHROME_PATH`
  • Training needs a separate environment pinned to `torch==2.9.1+cu129`

Getting started

The base package provides pixelshot, and extras add the indexing and serving stages. For the Claude Code plugin, the README recommends uv tool install so pixelshot ends up on your PATH.

pip install pixelrag
pip install 'pixelrag[index,pdf]'
pip install 'pixelrag[serve]'
uv tool install pixelrag

Requirements: Python 3.10+ (per the local PDF example); Chrome or Chromium, auto-installed only on linux-x64; poppler for PDF rendering; GPU optional (CUDA or MPS, with CPU fallback).

Always follow the official instructions in the README on GitHub. This overview is our summary, not a copy.

Sources