RAG-Anything is an all-in-one multimodal retrieval-augmented generation framework from the HKUDS lab, built on top of LightRAG. Where most RAG systems only handle plain text, RAG-Anything processes complete real-world documents: text, images, tables, charts and mathematical equations, and answers questions across all of them through a single pipeline. Released in 2025, it has quickly gathered about 24,000 stars on GitHub and is now integrated into LightRAG itself.
Modern documents are mixed-media: a financial report or research paper interleaves prose with figures, tables and formulas that text-only RAG silently drops. RAG-Anything addresses this with an end-to-end pipeline: documents in formats such as PDF and Office files are parsed with MinerU, Docling or PaddleOCR, content is routed to specialized analyzers for visual, tabular and mathematical material, and everything is indexed into a multimodal knowledge graph that captures cross-modal relationships. Queries then use modality-aware hybrid retrieval, combining vector search with graph traversal, and a VLM-enhanced query mode can reason over retrieved images together with text. The package installs from PyPI, supports batch processing and custom modality processors, and can also load an existing LightRAG instance.
RAG-Anything is free and open source under the MIT license. You run it on your own infrastructure; the only costs are compute and any external model APIs you plug in.
RAG-Anything is for researchers and developers building question-answering and knowledge systems over rich documents, such as academic papers, technical documentation, financial reports and enterprise knowledge bases, where the answers live in figures and tables as much as in text.