GraphRAG

GraphRAG is the retrieval-augmented generation pattern published by Microsoft Research in 2024 (Edge et al. "From Local to Global: A Graph RAG Approach to Query-Focused Summarization") that uses a knowledge graph as the retrieval substrate rather than (or in addition to) a vector database, enabling multi-hop reasoning, holistic question answering across an entire corpus, and structured citation that pure vector RAG struggles with. The Microsoft GraphRAG pipeline: (1) use an LLM to extract entities and relationships from source documents, producing a knowledge graph; (2) cluster the graph using hierarchical community detection (Leiden algorithm), producing summaries at multiple levels of abstraction; (3) at query time, route queries to either local search (graph traversal from extracted entities) for specific factual questions or global search (aggregation across community summaries) for thematic and holistic questions. The pattern decisively outperforms naive vector RAG on benchmarks involving multi-hop reasoning ("What did Alice's manager's company announce in 2023?"), holistic synthesis ("What are the main themes across these 500 reports?"), and structured exploration ("Show me everyone connected to this entity within two hops"). The cost is real: building the graph requires an LLM call per document to extract entities and relationships, and community summarization adds another layer of LLM calls. Microsoft's reference implementation (github.com/microsoft/graphrag) and a growing ecosystem (LightRAG, nano-graphrag, neo4j-graphrag, LlamaIndex KnowledgeGraphIndex, LangChain GraphCypherQAChain) make the pattern accessible. The hybrid approach — combining vector retrieval for similarity-based finding and graph retrieval for relationship-based finding — is the practical sweet spot for most production systems. AI governance teams favor GraphRAG for compliance-heavy applications because the graph provides structured, auditable citations rather than the diffuse passage-level citations of vector RAG.

GraphRAG on a 25-year-old taxonomy backbone: Centralpoint's 25 years of taxonomy, audience, and entity-relationship discipline means the graph substrate GraphRAG needs is already maintained in the platform — the AI layer composes against existing relationships rather than building them from scratch. GraphRAG runs on-premise, tokens meter per skill, and graph-grounded chatbots deploy through one line of JavaScript.


Related Keywords:
GraphRAG,GraphRAG,Oxcyon, AI, AI Governance, Generative AI, Inference, Inference, Inferencing, RAG, Prompts, Skills Manager,