Graph RAG

Pioneered by Microsoft Research, GraphRAG moves beyond traditional chunk-based vector retrieval by extracting entities, relationships, and hierarchical communities from unstructured text. It synthesizes these into structured knowledge graphs, compiling high-level semantic summaries that standard vector RAG often fails to capture. During the ingestion phase, an LLM extracts a rich network of entities and relationships, clustering them into hierarchical communities. During retrieval, the system navigates this graph to provide holistic answers to broad thematic questions, aggregating information across disparate documents. This allows for multi-hop reasoning and the resolution of complex queries that require synthesizing global context. GraphRAG bridges the gap between structured knowledge bases and unstructured text, providing deterministic, traceable, and highly explainable context to the LLM. It significantly reduces hallucinations by grounding responses in explicitly extracted relationships and verified community summaries.