What is GraphRAG?

GraphRAG is RAG that retrieves from a knowledge graph — a map of entities and how they relate to each other — instead of, or alongside, plain chunks of text.

Why does regular RAG struggle with some questions?

Regular RAG breaks your documents into chunks, turns them into vectors, and at query time fetches the chunks that look most similar to your question. That works fine when the answer lives in one paragraph. It falls apart on "multi-hop" questions, where the answer requires connecting facts scattered across several documents — say, "how are these two companies' executives connected?" No single chunk contains that whole story, so similarity search comes up empty or gives you a half-answer.

How does GraphRAG work?

GraphRAG fixes this by first extracting entities (people, companies, products) and their relationships into a graph, then letting the system traverse that graph — hopping from one connected fact to the next — instead of just matching text similarity. Because it follows real links between entities, it can assemble an answer that spans many documents. Microsoft popularized this approach with its open-source GraphRAG project, though the underlying idea (retrieve over a graph, not flat text) is a general pattern that other tools implement in their own ways too.

Here's how the two approaches compare:

When is GraphRAG worth the extra work?

The tradeoff is effort: building and maintaining that graph takes more upfront work than just chunking documents. You have to extract entities and relationships reliably and keep the graph current as your data changes. So GraphRAG tends to show up where data is deeply interconnected and the questions are genuinely multi-hop — legal discovery, corporate research, scientific literature — rather than as a default choice for every RAG project.

If your answers usually live in a single passage, plain chunk-based RAG is simpler and probably enough.