hkuds/rag-anything

RAG-Anything is an all-in-one Python framework for building question-answering systems that work with complex, mixed-content documents, not just plain text. RAG stands for Retrieval-Augmented Generation, a technique where an AI model answers questions by first searching a document collection for relevant information, then using that context to generate an answer. Most RAG systems struggle with documents that contain images, charts, tables, or mathematical equations alongside text. RAG-Anything is designed specifically to handle all of these content types together. The framework processes documents end-to-end: it ingests PDFs, Office files, and images, parses them into their component parts (text, tables, figures, equations), builds a multimodal knowledge graph that captures relationships between these elements, and then allows users to query across all of them through a single interface. It is built on top of LightRAG, another project from the same research group at Hong Kong University. A recent addition is VLM-Enhanced Query mode, which routes visual content through a vision-language model for deeper analysis when images are relevant to a query. This system is aimed at research and enterprise scenarios where documents contain rich mixed content, academic papers with figures and equations, financial reports with charts and tables, or technical documentation with diagrams. A Python package called raganything is available on PyPI, and the project has an accompanying academic paper on arXiv (2510.12323).