#2525th of 32 in RAG, knowledge and searchFor data and documents
pixelrag
Visual RAG that renders pages and PDFs to screenshots and searches the images
Live demo ↗
(opens in a new tab)Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 10.3k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Last release
- Jul 2026
- Language
- Python
1 on 26 May 202610,323 on 20 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
PixelRAG renders web pages, PDFs and images to screenshot tiles with Playwright/CDP, embeds them with a LoRA-fine-tuned Qwen3-VL-Embedding-2B, and serves FAISS (or Qdrant) search through a FastAPI endpoint. Tables, charts and layout are kept because nothing is parsed to text. It also ships a pixelbrowse plugin for Claude Code and opencode, and a hosted index of 8.28M Wikipedia pages.
Who it is for: Engineers building document search where tables, charts and layout matter
Strengths
- Retrieves over screenshots, so tables, charts and layout survive indexing
- Hosted API over 8.28M Wikipedia pages, no setup or API key
- Stages run standalone: pixelshot, chunk, embed, build-index, serve
- Runs on Linux CUDA and Apple Silicon MPS; FAISS default, optional Qdrant backend
Weaknesses
- Pre-built base Wikipedia index is about 217G to download
- Docker image is not built in CI and runs Chrome without its sandbox
- Training is a separate uv project with pinned torch 2.9.1+cu129
- Bundled turbo headless_shell auto-installs on linux-x64 only
What it needs
- GPU optional
- Docker + Compose
- Needs Chromium/Chrome (Playwright), FAISS, poppler (PDF), Qdrant (optional)
- Models: Qwen3-VL-Embedding-2B, LoRA adapters on Hugging Face
- port 30001
Also in RAG, knowledge and search
See all 32| Rank | Project | Score |
|---|---|---|
| 1 | deeptutorSelf-hosted AI tutor with knowledge bases, guided learning, books and quizzes | 82 out of 100 |
| 2 | LightRAGGraph-based RAG framework with a REST API server and web UI | 77 out of 100 |
| 3 | RAGFlowRAG engine with deep document parsing, template chunking and agentic retrieval | 74 out of 100 |
| #4 | MorphicAI answer engine with cited search results and generative UI components | 74 out of 100 |
| #5 | Open NotebookSelf-hosted NotebookLM alternative with podcasts and 20+ model providers | 73 out of 100 |