Glossary

GraphRAG

GraphRAG is retrieval-augmented generation that first turns documents into a knowledge graph of entities, such as people, systems and companies, and the relationships between them. Retrieval then follows those links and community summaries, which suits questions whose answer spans many documents or needs a big-picture view.

Which questions does vector retrieval miss?

Standard RAG returns the passages whose embeddings sit closest to the question. That works when one passage holds the answer. It struggles with a question such as which suppliers connect two incidents, because no single passage says so. The answer exists only as a connection across ten documents, and similarity to the question wording does not reveal connections. A second kind of miss is the broad question, such as what the main themes are across a thousand reports, where no handful of passages can represent the whole.

How does GraphRAG work?

The method was described by Microsoft Research in 2024. At indexing time a language model reads every passage and extracts entities and the links between them. Closely connected entities are grouped into communities, and a summary is written for each community. At question time the system follows links from the entities in the question and reads the relevant community summaries, then passes that evidence to the model. Facts that were scattered in the text become traversable, and the summaries give an answer to big-picture questions that no individual passage could.

What does GraphRAG cost?

Indexing runs a model over the whole corpus, not just an embedding model over it, so building a graph takes far more compute than building a vector index, often on accelerators (see GPU storage). Maintenance adds work too. The same company appears as "IBM" in one document and "I.B.M." in another and has to be merged into one node. When a source document is deleted, the entities and links drawn from it have to be removed or the graph keeps facts that no document supports.

The graph files are derived output. They can be regenerated from the sources, which makes them a different class of data from the original documents and often a candidate for lighter protection.

When is plain vector retrieval enough?

If most questions are answered from a single passage, a vector index alone is cheaper and sufficient. Teams that need both often run them side by side, using vector or hybrid search for direct lookups and the graph for connected questions. It is one of two ways to handle multi-step questions. The other is agentic RAG, which repeats searches at question time instead of connecting facts in advance.

Frequently asked questions

Is GraphRAG the same as a graph database?

No. A graph database is a storage engine for nodes and edges. GraphRAG is a retrieval method that builds such a graph from unstructured documents using a language model, and it can store the result in several ways.

Does GraphRAG replace a vector index?

Usually not. Vector search remains the efficient route for direct lookups, and many systems keep both.

Where does the graph data live?

Source documents and the generated graph files are separate layers, since the sources are irreplaceable and the graph output can be rebuilt. This split is described under RAG storage.