LightRAG overview and key facts (October 2026)

LightRAG is an open-source RAG system that extracts entities and relationships from documents into a knowledge graph and uses both the graph and vector search to answer questions. It targets developers who need answers that connect facts across many documents.

LightRAG comes from the Data Intelligence Lab at the University of Hong Kong and was presented at EMNLP 2025. It is published under the MIT license and had about 39,900 GitHub stars as of September 28, 2026, with commits on that day. Storage backends include PostgreSQL, MongoDB, Neo4j and OpenSearch. A web interface lets users insert documents, query them and view the graph. Recent releases added four chunking strategies, separate model settings for extraction and queries, document deletion with graph regeneration, and evaluation with RAGAS plus tracing with Langfuse.

• Extracts a knowledge graphA language model reads each document and records entities and their relationships in a graph. Teams get a structured map of people, products or terms that appear across their files.
• Answers questions across documentsQueries combine graph traversal with vector search, and a reranker sorts the results for mixed queries. Questions that need facts from several documents get more complete answers than with pure vector search.
• Handles tables, images and formulasSince the merge of RAG-Anything in May 2026, multimodal content can be parsed through MinerU or Docling services. Reports with tables and figures become part of the searchable knowledge base.
• Cites its sourcesAnswers can include citations, and the API returns the retrieved context. Reviewers can trace a statement back to the document and measure retrieval quality.

• Uses a knowledge graph as a core part of retrieval, which helps with questions about relationships between entities.
• Allows different language models for extraction, queries, keywords and image understanding, so costs can be tuned per step.
• Maintained by a university lab under the MIT license, without a commercial edition.

Research project of the Data Intelligence Lab at the University of Hong Kong without a country focus, usable worldwide as open source.

Building the graph sends every document through a language model, so indexing costs grow with the volume of text; test on a sample first. Choose a model provider whose data terms fit your documents, or run open models locally. There is no vendor support contract, so plan who maintains the installation and its storage backends.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews