Repolode
Submit
HKUDS

HKUDS/RAG-Anything

All-in-one RAG framework for querying documents that mix text with images, tables, and equations, built on LightRAG.

View on GitHubWebsite
Stars
23,899
Forks
2,787
Open issues
123
Last commit
8 days ago

What is it?

RAG-Anything is a retrieval-augmented generation framework for documents that contain more than plain text. Standard RAG pipelines struggle with figures, tables, and mathematical formulas, so this project handles them in a single pipeline alongside the surrounding text.

It parses PDFs and Office files (DOC, DOCX, PPT, PPTX, XLS, XLSX) through MinerU, with Docling and PaddleOCR also selectable as parsers. Images, tables, and equations are routed to dedicated processors that generate descriptions and entities, which are then added to a knowledge graph built on LightRAG. Queries run over that graph with hybrid retrieval, and a vision model can be used to analyze images found in the retrieved context.

The project provides a Python API for end-to-end document processing, batch folder processing, direct insertion of pre-parsed content lists, and multimodal queries that accept tables or LaTeX equations as input. It is published on PyPI as `raganything`.

Key features

  • 01Parses PDFs and Office documents into text, images, tables, and equations.
  • 02Routes multimodal elements to dedicated processors for captioning and analysis.
  • 03Builds a knowledge graph with cross-modal relationships on top of LightRAG.
  • 04Supports hybrid, local, global, and naive query modes for text queries.
  • 05Accepts pre-parsed content lists so parsing can be skipped.
  • 06Offers batch processing for whole folders of documents.

Use cases

  • Question answering over research papers that rely on figures and formulas.
  • Searching financial or technical reports whose key data sits in tables.
  • Building enterprise knowledge bases from mixed-format office documents.
  • Integrating pre-parsed document output from external tools into a RAG index.

Things to know

  • Document parsing relies on the MinerU project, which adds a dependency to set up.
  • Image analysis expects a vision model, so you will need access to one.

Topics

#rag#multimodal#retrieval-augmented-generation#knowledge-graph#document-parsing#llm#pdf#multi-modal-rag

FAQ

What is RAG-Anything?
All-in-one RAG framework for querying documents that mix text with images, tables, and equations, built on LightRAG. RAG-Anything is a retrieval-augmented generation framework for documents that contain more than plain text. Standard RAG pipelines struggle with figures, tables, and mathematical formulas, so this project handles them in a single pipeline alongside the surrounding text.
Is RAG-Anything free to use? What license does it have?
RAG-Anything is open source under the MIT license. Check the license terms before using it in commercial products.
Is RAG-Anything still maintained?
Actively maintained. The last commit to RAG-Anything was on Oct 12, 2026.
What can I use RAG-Anything for?
Question answering over research papers that rely on figures and formulas. Searching financial or technical reports whose key data sits in tables. Building enterprise knowledge bases from mixed-format office documents. Integrating pre-parsed document output from external tools into a RAG index.
Where can I get RAG-Anything?
The source code and installation instructions are on GitHub: https://github.com/HKUDS/RAG-Anything

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects