GraphRAG

A family of RAG architectures that uses a graph of entities, relationships and, in some designs, communities or summaries to retrieve structurally connected context. It helps when a question depends on relationships across documents, multi-hop reasoning, or a corpus-wide view, situations in which isolated chunks lose the structure of the problem.

GraphRAG is not one algorithm and not an automatic replacement for vector search. Extracting, updating, and evaluating the graph is costly and introduces new failure points: incorrect entities, invented relationships, low coverage, and weak provenance. It is justified only when relational structure improves a relevant metric against a simpler baseline.