StarTrail-org

PixelRAG: Search Documents by Their Visual Content

PixelRAG indexes screenshots of web pages and documents for visual retrieval, preserving charts, tables, and layout that text extraction can lose. It offers a Python pipeline, command-line renderer, and hosted search API.

PythonApache-2.0AIRAGComputer Vision
View on GitHub Compare
Stars
10.1k
Forks
870
Stars, last 30 days
+659

Our take

Health score

54/100

How it is scored
Activity
57
Community
63
Issues
53
Pull requests
42

+659 stars in the last 30 days

  • Slower to merge pull requests than 89% of the projects we track

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

10,965

Forks

987

Change since Nov 17, 2026: +12 stars·0 forks

Overview

PixelRAG is a visual retrieval system that renders web pages, PDFs, and other documents as images, then retrieves relevant screenshot tiles. It targets a weakness of text-based retrieval: parsing can discard meaningful visual structure such as charts, tables, and page layout.

It is intended for developers building retrieval-augmented generation or document search where visual details matter. It also provides a renderer that AI coding agents can use to inspect screenshots instead of relying only on raw HTML.

Key Features

  • Render web pages and PDFs into screenshot tiles with the pixelshot CLI or Python API.
  • Embed and retrieve visual tiles using Qwen3-VL-Embedding, with a LoRA-fine-tuned model available.
  • Build and serve local FAISS indexes, or use Qdrant as an index backend.
  • Search a hosted index of Wikipedia pages through the PixelRAG API, including image queries.
  • Run indexing as separate stages or orchestrate ingestion, embedding, and indexing with pixelrag index.
  • Integrate screenshot-based browsing into Claude Code and opencode workflows.

Use Cases

  • RAG developers can index documents where answers depend on tables, diagrams, charts, or page composition.
  • Researchers can search collections of papers and PDFs by the visual information on their pages.
  • Teams can build local document-search services with FAISS, or use Qdrant for configurable, disk-backed vector storage.
  • Agent developers can let Claude Code or opencode inspect rendered web pages when HTML parsing misses important details.

Getting Started

Install the package and render a page:

pip install pixelrag
pixelshot https://en.wikipedia.org/wiki/Python --output ./tiles

For indexing, serving, and optional integrations, see the README. A hosted search API and browser demo are available at pixelrag.ai.

Alternatives

  • PaddleOCR: PaddleOCR parses text and document layouts from images and PDFs, while PixelRAG indexes page screenshots for visual retrieval.
  • xberg: Xberg extracts text, tables, and structured data from many file types, while PixelRAG retrieves from rendered page images.
ProjectLanguageLicenseStarsStatus
PixelRAGPythonApache-2.010.1kNot checked yet
PaddleOCRPythonApache-2.090.6kActive
xbergRustMIT9.4kActive

Considerations

  • Visual indexing requires rendering and image embedding, so it involves more processing than text-only indexing. The README describes CUDA, Apple Silicon/MPS, and CPU fallback options; performance will depend on the available hardware.
  • PDF rendering requires the pdf extra and Poppler. Web rendering relies on Chrome or Chromium, with automatic bundled browser installation limited to Linux x64.
  • A pre-built Wikipedia index is available through the hosted API. Self-hosting the documented base index requires downloading a large index, listed as approximately 217G in the README.
  • Training is a separate project with its own pinned environment. Most users can use the published adapters without retraining.

Found this useful?

Share it with someone who would like PixelRAG.

Comparisons