Retrieval-augmented generation where the external knowledge source is graph-structured, so retrieval must exploit relational structure rather than only embedding similarity. GraphRAG couples graph machine learning (e.g., graph neural networks) and graph/network analysis (traversal search, community detection) with dense/embedding retrieval to fetch relevant nodes, edges, subgraphs, or relational paths for a language-model generator. This survey organizes the space with a holistic framework of five components, query processor, retriever, organizer, generator, and data source, and reviews domain-tailored designs across knowledge, document, molecule, and social/product graphs.
A survey of GraphRAG: retrieval-augmented generation in which the external knowledge is graph-structured (knowledge graphs, document graphs, molecule graphs, social/product networks). Whereas convent…