Running GraphRAG Locally with LightRAG and RAG-Anything

Setting up a GraphRAG system in your local environment does not need to be complicated. Two open-source GitHub repositories, used in tandem, give you a fully functional GraphRAG pipeline that handles both text and multimodal documents. The key is understanding what each one does and how they fit together.

LightRAG: The Foundation

LightRAG is the core engine. It lays the foundation for your GraphRAG system by incorporating graph structures into text indexing and retrieval. It uses a dual-level retrieval system -- low-level and high-level knowledge discovery -- that consistently outperforms competing approaches like NaiveRAG, HyDE, and even Microsoft's GraphRAG across multiple domains.

LightRAG is completely open source. The only limitation is that it can only ingest text documents out of the box. No images, no charts, no graphs -- text only.

RAG-Anything: The Multimodal Wrapper

RAG-Anything solves LightRAG's text-only limitation. Built by the same team at HKU, it wraps around LightRAG and adds multimodal document processing. Non-text documents get routed through a processing tunnel that extracts text content and pushes images through OCR.

The result is a GraphRAG system that gets the best of both worlds: graph-structured retrieval from LightRAG with multimodal ingestion from RAG-Anything.

Component	Role	Handles
LightRAG	Core GraphRAG engine	Text documents only
RAG-Anything	Multimodal wrapper around LightRAG	Images, charts, PDFs, mixed documents
Combined	Full GraphRAG pipeline	Everything, completely open source

Key Takeaways

• Use LightRAG and RAG-Anything together for a complete local GraphRAG setup
• LightRAG handles text indexing with graph-structured retrieval
• RAG-Anything wraps LightRAG to add multimodal document support (images, charts, OCR)
• Both are open source and built by the same team at HKU
• The combined stack runs entirely in your local environment

Resources

• LightRAG on GitHub -- The core GraphRAG engine (EMNLP 2025)
• RAG-Anything on GitHub -- Multimodal wrapper for LightRAG
• LightRAG Paper -- "LightRAG: Simple and Fast Retrieval-Augmented Generation"