GraphRAG, published by Darren Edge, Ha Trinh, and colleagues at Microsoft Research in April 2024, addresses a blind spot in ordinary retrieval-augmented generation. Standard RAG fetches a few passages most similar to a query, which works for narrow lookups but fails on global questions aimed at an entire corpus, such as “what are the main themes across all these documents.” Those are really summarization tasks, and similarity search alone cannot gather the whole picture.
GraphRAG works in stages. An LLM first reads the source documents and extracts an entity knowledge graph linking people, places, and concepts. It then groups closely related entities into communities and pre-generates a summary for each community. To answer a broad query, the system produces partial answers from the relevant community summaries and synthesizes them into a final response.
On global sensemaking queries over corpora up to a million tokens, the authors reported substantial improvements over conventional RAG in both the comprehensiveness and the diversity of answers. Microsoft released GraphRAG as an open-source toolkit, and it became a widely cited approach for corpus-level question answering.
For a business, GraphRAG is the technique to reach for when users ask sweeping questions about a large document collection rather than looking up a single fact, the kind of query where plain vector search comes back thin.