GraphRAG

GraphRAG models knowledge as an entity and relationship graph (subject-relation-object triples). Two key advantages: multi-hop reasoning over relations (user → doctor → hospital → address) and entity disambiguation (two “Dr. Zhangs” are different graph nodes). A community detection algorithm identifies thematic clusters with summaries. Limitation: Triple decomposition loses conditional logic and temporal dependencies of natural language. Recommendation: store full text as is, graph as a specialized index.

Related: RAPTOR, [Structured Indexing], [User Memory Formats]