What's the difference between RAG and GraphRAG?

Standard RAG retrieves the chunks closest to a question and answers from them; GraphRAG, from Microsoft Research, first builds a knowledge graph of entities out of the whole corpus, clusters it into communities, and summarizes each one, so it can also answer questions no single chunk covers.

Microsoft’s own framing is that “baseline RAG struggles to connect the dots” when a question needs traversing several pieces of information through their shared attributes, and “performs poorly when being asked to holistically understand summarized semantic concepts over large data collections.” Top-k retrieval only ever hands the model a handful of chunks, so a question like “what are the main themes across this dataset” has no single chunk that contains the answer. GraphRAG’s community summaries exist for that: instead of retrieving passages, it draws partial answers from the summaries and combines them.

That coverage costs something standard RAG doesn’t pay: the graph and its summaries are built once, over the whole corpus, before any question is asked. It earns that cost on aggregation questions, where a chunk-based answer that reads as grounded can still be wrong. For a question a handful of chunks already answers, the graph buys nothing standard RAG didn’t have.