GraphRAG beats standard vector RAG by 10 to 13 accuracy points on complex reasoning and contextual summarization tasks, and wins 72 to 83 percent of comprehensiveness comparisons on global corpus questions. That is the finding across the original Microsoft paper and four independent benchmark studies reviewed here. For simple fact retrieval, text chunks and graphs tie at roughly 60 points each. The graph's edge grows directly with the reasoning depth of the question.
The mechanism matters. Standard vector RAG retrieves the k most similar passages and fails structurally on multi-hop questions because no single chunk contains the answer. GraphRAG builds a knowledge graph during indexing using an LLM to extract entities and relationships, runs the Leiden algorithm for community detection, pre-writes summaries per cluster, and at query time uses a map-reduce synthesis over those summaries. On multi-hop benchmarks MuSiQue, HotpotQA, and 2WikiMultiHopQA, graph-guided retrieval lifts Recall@5 from 73.4 percent to 87.8 percent, with MuSiQue alone jumping 31 points. HippoRAG, which uses Personalized PageRank instead of community summaries, reports 20 percent accuracy gains at 10 to 20 times lower cost than iterative retrieval methods.
The full article is worth reading for two reasons the headline obscures. First, a 2025 Michigan State and Meta study using a unified evaluation protocol found plain RAG still edges ahead on single-hop factual lookup, F1 64.8 versus 63.0. The graph is a specialized tool, not a universal upgrade. Second, index construction costs roughly 48 dollars against GPT-4o for a moderate corpus, and many benchmark wins are judged by LLMs that independent research shows are biased toward longer, more elaborated answers, which graphs tend to produce. Microsoft's own LazyGraphRAG cuts that index cost to 0.1 percent of the original by deferring extraction to query time, a tacit admission the upfront budget is impractical for most deployments.
[READ ORIGINAL →]