StarTrail-org

StarTrail-org / PixelRAG

Public

The end of web parsing. The beginning of scalable pixel-native search.

agent ai memory multimodal rag
11
0
85% credibility
Found May 29, 2026 at 24 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
JSON
AI Summary

PixelRAG is a visual search system that lets you turn any collection of documents into a searchable image database. Instead of extracting text (which loses formatting), it preserves documents as screenshots and uses AI to find the most relevant images for your questions. You can build search indexes from Wikipedia, local files, PDFs, or web pages, then query them with natural language questions. The system returns the exact screenshot tiles containing the answer, so you see the visual context rather than just extracted text.

How It Works

1
📚 You have a collection of documents

You gather your PDFs, web pages, and local files that you want to search through visually.

2
🔨 You turn them into pictures

The system takes each page and renders it as a screenshot image, preserving the exact layout and formatting.

3
💡 You ask a question in plain English

You type a question like 'What is the capital of France?' and the system searches through your screenshot images to find the most relevant ones.

4
📈 You receive visual answers

The system shows you the exact screenshot tiles that contain the answer, with relevance scores so you can see why each result matched.

✨ You get accurate answers with visual proof

You see the actual document pages with the information, complete with tables, images, and formatting preserved exactly as in the original.

Sign up to see the full architecture

3 more

Sign Up Free

Star Growth

See how this repo grew from 24 to 11 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is PixelRAG?

PixelRAG is a visual search framework that captures documents as screenshots instead of parsing text, then uses vision-language models to build searchable vector indexes. It works with web pages, PDFs, and local files, rendering everything as image tiles that get embedded and indexed with FAISS. The system ships as five independent packages: render (document to images), embed (images to vectors), index (orchestration pipeline), serve (FastAPI search API), and train (LoRA/DoRA fine-tuning for Qwen3-VL). You can download pre-built Wikipedia indexes and query them immediately, or build custom indexes from your own documents via a YAML config file.

Why is it gaining traction?

The pitch is simple: instead of fighting with HTML parsers, PDF extractors, and table detection, just screenshot everything. Layout, tables, figures, and formatting are preserved automatically because the model sees what humans see. The project includes a Claude Code plugin that gives Claude "eyes" to screenshot URLs directly from conversations, with no MCP server required. Pre-built Wikipedia indexes with millions of articles are available for download, so you can run semantic visual search without processing anything yourself.

Who should use this?

RAG developers struggling with extraction failures on complex layouts should try this as an alternative approach. Agents building multi-step research tools will benefit from the Claude plugin integration. Teams working with heterogeneous document collections (web pages mixed with PDFs) where traditional parsing breaks down will find the unified screenshot pipeline valuable. Researchers benchmarking visual retrieval methods have a complete end-to-end system here, not just a library. This is not yet ready for production if you need stability guarantees or comprehensive documentation.

Verdict

PixelRAG tackles a real problem with a clever idea, but 11 stars and a 0.85% credibility score signal a project in early exploration rather than production-ready infrastructure. The code is well-structured and the demos work, but the maturity markers (limited test coverage, sparse docs beyond README, no enterprise support) make it a strong candidate for prototyping and experimentation rather than mission-critical pipelines. Try it on a contained use case first.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.