HKUDS/LightRAG

[EMNLP2025] "LightRAG: Simple and Fast Retrieval-Augmented Generation"

What it solves

LightRAG is a lightweight, graph-based retrieval-augmented generation (RAG) framework designed to be a more efficient alternative to Microsoft GraphRAG. It solves the problem of fragmented context in traditional chunk-based RAG and the high computational cost, slow response times, and expensive incremental updates associated with large-scale graph indexing.

How it works

LightRAG uses a dual-layer architecture that manages both knowledge graphs (KGs) and vector embeddings. This allows it to bridge the gap between vector-based and graph-based RAG. It employs a dual-level retrieval mechanism to integrate detailed facts and abstract concepts simultaneously. For updates, it uses a seamless incremental indexing pipeline where new data is merged into the existing graph without rebuilding the global index.

Who it’s for

It is designed for users needing high-quality RAG in vertical domains (such as legal or financial) that require global comprehension or logical reasoning, as well as those working with dynamic datasets that require real-time updates.

Highlights

  • Dual-Layer Architecture: Combines knowledge graphs and vector embeddings for deep contextual understanding.
  • Incremental Updates: Supports seamless, real-time knowledge base updates and efficient document deletion.
  • Multimodal Support: Integrates with RAG-Anything to parse and retrieve text, tables, formulas, and images from documents.
  • Role-Specific LLM Config: Allows independent LLM settings for extraction, querying, keywords, and vision tasks to balance cost and performance.
  • Flexible Storage: Supports multiple backends including MongoDB, PostgreSQL, Neo4J, and OpenSearch.