A review of evidence from Microsoft’s original paper and four independent benchmark studies shows that GraphRAG substantially improves retrieval-augmented generation (RAG) answers, but only for specific question types and at a cost. Standard vector RAG, which retrieves the k passages most similar to a query, has three structural blind spots: it cannot connect facts across different passages, it is blind to global questions like “What are the main themes?” because similarity search returns only a handful of superficially similar chunks, and it severs context at chunk boundaries. Microsoft Research noted that baseline RAG “struggles to connect the dots” and performs poorly on “holistically understand summarized semantic concepts over large data collections.” GraphRAG addresses this by having a large language model extract entities, relationships, and claims from every chunk during indexing, assembling them into a weighted knowledge graph. It then runs community detection using the Leiden algorithm to cluster the graph into a hierarchy, pre-writing natural-language summaries for each community. At query time, relevant communities draft partial answers, which are ranked and merged before the model synthesizes a final response. Variants like HippoRAG use the graph plus a Personalized PageRank walk instead. The evidence pattern shows the headline win is global sense making.