GraphRAG
also called Graph retrieval
Retrieval over an entity graph, for questions about the whole corpus.
An approach that builds a graph of entities and relationships with community summaries, so questions like "what are the themes across these documents" become answerable. It complements chunk retrieval rather than replacing it.
The gap it fills is structural: similarity search returns the few most relevant passages, which is useless when the answer requires synthesising a hundred documents. Building the graph is expensive, so it earns its place on corpora that get asked aggregate questions repeatedly.
The cost is front-loaded: extracting entities and relationships and summarising communities is an indexing job that runs over the whole corpus, and it has to be rerun as the corpus changes. That is the reason it is worth it only where aggregate questions recur.
The question chunk retrieval cannot answer
A stakeholder asks what themes emerge across two years of incident reports.
Any set of retrieved chunks is a sample and the answer will be an artefact of which ones were selected. Community summaries over the whole corpus produce an answer that is actually about the corpus rather than about your retrieval settings.
- paperFrom local to global: a graph RAG approach to query-focused summarization ↗
Build an entity graph with community summaries so "what are the themes across this corpus?" becomes answerable — the question chunk retrieval structurally cannot answer.
- reportProject GraphRAG ↗
Microsoft’s own project page and publication list for GraphRAG, useful as the publisher-side companion to the paper.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Knowledge Cutoff
The date the model’s knowledge stops.
Hybrid Search
Keyword search and vector search, combined and merged.
Reranking · Reranker
A second pass that reorders candidates by actual relevance.
Chunking
Splitting documents into retrievable pieces.