GraphRAG
A RAG approach that builds knowledge graphs from documents to enable more structured and comprehensive retrieval.
Overview
GraphRAG extends traditional vector-based RAG by constructing knowledge graphs from source documents. Entities and relationships are extracted and organized into a graph structure, enabling more structured and comprehensive retrieval than pure vector similarity search.
Advantages
GraphRAG excels at answering questions that require synthesizing information across multiple documents, understanding relationships between entities, and providing comprehensive summaries of a topic. Microsoft's GraphRAG implementation uses LLMs to extract entities and relationships, build community summaries at multiple levels, and combine graph-based retrieval with traditional vector search for more complete answers.