GraphRAG is a variant of retrieval-augmented generation that combines a knowledge graph with vector retrieval, using the explicit relationships in the graph to provide richer, more structured context to a language model than vector search alone can.
Plain RAG retrieves chunks that are individually similar to a query but unaware of how they relate to each other. GraphRAG adds a layer of structure: entities and the relationships between them are organised into a graph, so the system can follow connections — this company acquired that one, which makes this product, used by these customers — and assemble context that spans multiple linked facts. This is especially valuable for multi-hop questions that require connecting several pieces of information.
In practice GraphRAG often blends the two retrieval modes: vector search finds relevant entry points, and graph traversal expands outward to gather related entities and facts. The result is context that captures not just topically similar text but the reasoning structure connecting it, improving answers on complex, relationship-heavy questions at the cost of building and maintaining the knowledge graph.