Reference: https://github.com/HKUDS/RAG-Anything

 

What it is

 

An all-in-one multimodal RAG framework built on top of LightRAG (also from HKUDS) that handles text, images, tables, and mathematical equations within a single unified pipeline — instead of needing separate tools for each content type. 15.1k stars, actively maintained, with a technical report on arXiv.

 

Use case

 

For documents that aren’t just plain text — research papers, financial reports, technical docs with charts/tables/formulas — RAG-Anything parses them via MinerU/Docling/PaddleOCR, routes each content type (text, image, table, equation) through specialized processors, and builds a multimodal knowledge graph that captures cross-modal relationships (e.g., a table linked to the paragraph that references it). You can then query with pure text, VLM-enhanced queries (where retrieved images are sent directly to a vision model), or multimodal queries (pass a specific table/equation alongside your question). It also supports direct content-list insertion if you already have pre-parsed output from another tool.

 

Benefit

 

• One pipeline instead of five — eliminates stitching together separate OCR, table-extraction, and image-captioning tools before you even get to retrieval
• Relationship-aware retrieval — combines vector similarity with knowledge-graph traversal, so retrieval understands that a chart belongs to a specific section rather than treating every chunk as flat and independent
• VLM-in-the-loop — when a query touches on visual content, it automatically pulls in the actual images for the vision model to reason over, not just an image caption as a text proxy
• Flexible entry point — process_document_complete() for raw files, or insert_content_list() if you’re feeding in output from something like ade-python/LandingAI, which means it could plug in downstream of the LandingAI tools you just covered

 

Happy Learning!!