Graph Retrieval-Augmented Generation
Also known as GraphRAG
Overview
Graph Retrieval-Augmented Generation studies how to use entities, relations, and document structure to help language models find and combine relevant evidence. Systems build knowledge graphs or hypergraphs from documents, then retrieve evidence through paths, graph traversal, community summaries, or hybrid graph-and-vector search; some use agents to plan and verify multi-step retrieval. The methods are tested on knowledge-graph and multi-hop question answering, long-document tasks, and specialized settings such as medicine, law, and engineering, where traceable evidence can improve factuality. A central tension is whether richer graph structure improves answers enough to justify its construction and retrieval costs: performance depends on graph quality and query type, and more retrieved context does not always yield better answers.
Latest papers
232 papers19 papers in the last four weeks, down 21% on the four weeks before. 0.1% of all new papers.