RAG-Anything: Build Multimodal Document RAG

Overview

RAG-Anything is a Python framework for retrieval-augmented generation over documents that mix text with images, tables, equations, and other content. Built on LightRAG, it aims to keep those elements and their context available to the same retrieval and question-answering workflow instead of treating a document as plain text alone.

It is aimed at developers integrating multimodal document ingestion and retrieval into applications or internal knowledge bases. The project supports parser-based document processing as well as insertion of content lists prepared by other tools, giving teams a choice between using its pipeline and connecting an existing parser.

Key Features

• Processes PDFs, Office documents, images, and supported text files through configurable parsers.
• Handles text, images, tables, and equations, with modality-specific processing and model functions.
• Builds multimodal entities and relationships on top of LightRAG for retrieval that can combine vector search and graph context.
• Offers text queries, queries with supplied multimodal content, and optional vision-model analysis of retrieved images.
• Accepts pre-parsed content lists, so ingestion can connect to external parsing workflows.
• Supports folder and batch processing, custom modality processors, and parsers including MinerU, Docling, and PaddleOCR.
• Includes optional audio and video processing, with additional dependencies and configuration.

Use Cases

• Research groups can query papers containing figures, tables, and formulas without separating those elements from their surrounding text.
• Documentation teams can build a searchable knowledge base from PDFs and Office files with mixed layouts.
• Developers with an existing extraction pipeline can pass its parsed content lists into the framework rather than replace the parser.
• Organizations handling reports or presentations can ask questions that draw on both written passages and visual material, if they configure a compatible vision model.

Getting Started

Install the package:

pip install raganything

The README covers optional dependencies, model configuration, parser setup, and end-to-end examples: RAG-Anything README.

Alternatives

• kotaemon: kotaemon centers on a customizable document-QA interface and toolkit, rather than a multimodal ingestion workflow built on LightRAG.
• KAG: KAG emphasizes knowledge graphs and multi-hop reasoning over source text, rather than unified retrieval across text, images, tables, and equations.
• graphrag: GraphRAG builds graph-based context from unstructured text for question answering; it does not focus on parsing mixed-format documents.

Considerations

The framework is active, with regular commits and releases, and its contributor activity is spread across many people rather than dominated by one. However, multimodal ingestion still has operational requirements: Office document parsing requires LibreOffice, some file formats need optional Python dependencies, PaddleOCR needs a separate PaddlePaddle installation, and video processing requires ffmpeg. GPU use is supported for some parsing paths but is not stated as a universal requirement. Full RAG use also requires configuring language and, for image analysis, vision model functions. The README documents parser-specific gaps, including that public asset URL mapping is currently wired only into the MinerU path. Review the README and dependency setup against your formats and deployment environment before adopting it.