{ "post": { "title": "GraphRAG vs RAG: what the evidence actually says", "url": "https://bardak.ai/posts/graphrag-vs-rag-evidence", "updated": "2026-10-29T18:55:51.676Z" }, "citation": "GraphRAG vs RAG: what the evidence actually says, Ivan Bardak, bardak.ai, https://bardak.ai/posts/graphrag-vs-rag-evidence", "datasets": [ { "key": "three-studies-three-angles", "title": "Three studies, three angles", "type": "benchmarkTable", "source": "Edge et al. 2024; Han et al. 2025; Xiang et al. 2025", "caption": "", "csv": "https://bardak.ai/data/graphrag-vs-rag-evidence/three-studies-three-angles.csv", "header": [ "Study", "Setting", "Main finding" ], "rows": [ [ "Edge et al. 2024 (Microsoft Research, the GraphRAG authors)", "Whole-corpus questions over ~1M-token corpora; LLM judge, pairwise", "Global, summary-based answers were more comprehensive and diverse; vector RAG was more direct" ], [ "Han et al. 2025 (academic team)", "Single-hop, multi-hop and summary questions; scored against ground truth", "RAG wins single-hop, GraphRAG local wins multi-hop, combining both is best overall" ], [ "Xiang et al. 2025, GraphRAG-Bench (academic team)", "Fact retrieval to creative writing on a novel and a medical corpus; ground truth", "RAG equal or better on facts; graph methods ahead on reasoning and summaries, but not everywhere" ] ] }, { "key": "how-often-global-answers-beat-vector-rag-llm-judged-win-rate", "title": "How often global answers beat vector RAG (LLM-judged win rate)", "type": "chart", "source": "Edge et al. 2024, From Local to Global (Microsoft Research)", "caption": "50% would be a tie. Vector RAG still gave the most direct answers in every comparison, which the authors used as a validity check.", "csv": "https://bardak.ai/data/graphrag-vs-rag-evidence/how-often-global-answers-beat-vector-rag-llm-judged-win-rate.csv", "header": [ "measure", "Lowest of the global approaches", "Highest of the global approaches" ], "rows": [ [ "Comprehensiveness: podcasts", "72", "83" ], [ "Comprehensiveness: news", "72", "80" ], [ "Diversity: podcasts", "75", "82" ], [ "Diversity: news", "62", "71" ] ] }, { "key": "vector-rag-graphrag-local-and-both-combined-llama-3-1-70b", "title": "Vector RAG, GraphRAG local and both combined (Llama 3.1 70B)", "type": "chart", "source": "Han et al. 2025, RAG vs. GraphRAG: A Systematic Evaluation and Key Insights, tables for NQ, HotpotQA and MultiHop-RAG", "caption": "Single-hop favours plain RAG; multi-hop favours the graph; feeding the model both retrievals beat either one.", "csv": "https://bardak.ai/data/graphrag-vs-rag-evidence/vector-rag-graphrag-local-and-both-combined-llama-3-1-70b.csv", "header": [ "benchmark", "Vector RAG", "GraphRAG local", "RAG + GraphRAG local" ], "rows": [ [ "NQ: single-hop (F1)", "68.18", "65.44", "69.75" ], [ "HotpotQA: multi-hop (F1)", "63.88", "64.60", "66.50" ], [ "MultiHop-RAG: multi-hop (accuracy)", "65.77", "71.17", "77.62" ] ] }, { "key": "multihop-rag-accuracy-by-question-type-llama-3-1-70b", "title": "MultiHop-RAG accuracy by question type (Llama 3.1 70B)", "type": "chart", "source": "Han et al. 2025, MultiHop-RAG results", "caption": "Graphs help most on temporal and comparison questions. On questions with no answer in the corpus, global search answered anyway instead of saying it did not know.", "csv": "https://bardak.ai/data/graphrag-vs-rag-evidence/multihop-rag-accuracy-by-question-type-llama-3-1-70b.csv", "header": [ "question type", "Vector RAG", "GraphRAG local", "GraphRAG global" ], "rows": [ [ "Inference", "94.85", "92.03", "89.09" ], [ "Comparison", "56.31", "60.16", "66.00" ], [ "Temporal", "25.73", "49.06", "59.18" ], [ "Null (no answer in the corpus)", "91.36", "88.70", "13.95" ] ] }, { "key": "accuracy-by-task-novel-corpus-gpt-4o-mini", "title": "Accuracy by task, novel corpus (GPT-4o-mini)", "type": "chart", "source": "Xiang et al. 2025, When to use Graphs in RAG (GraphRAG-Bench), generation results", "caption": "On facts, vector RAG with a reranker matches the best graph method and beats Microsoft's GraphRAG. On reasoning and summaries, graph methods lead.", "csv": "https://bardak.ai/data/graphrag-vs-rag-evidence/accuracy-by-task-novel-corpus-gpt-4o-mini.csv", "header": [ "task", "Vector RAG with reranker", "MS-GraphRAG", "HippoRAG2" ], "rows": [ [ "Fact retrieval", "60.92", "49.29", "60.14" ], [ "Complex reasoning", "42.93", "50.93", "53.38" ], [ "Contextual summary", "51.30", "64.40", "64.10" ], [ "Creative generation", "38.26", "39.10", "48.28" ] ] }, { "key": "reasoning-accuracy-against-prompt-tokens-per-question-novel-corpus", "title": "Reasoning accuracy against prompt tokens per question (novel corpus)", "type": "chart", "source": "Xiang et al. 2025, generation results and average tokens per query", "caption": "More tokens did not buy more accuracy. HippoRAG2 reached the best reasoning score at almost the cost of plain vector RAG. Microsoft's GraphRAG is left out: the paper's text and table give different token counts for it.", "csv": "https://bardak.ai/data/graphrag-vs-rag-evidence/reasoning-accuracy-against-prompt-tokens-per-question-novel-corpus.csv", "header": [ "method", "tokens", "accuracy", "type" ], "rows": [ [ "Vector RAG", "879", "41.35", "Vector RAG" ], [ "HippoRAG2", "1008", "53.38", "Graph or tree RAG" ], [ "RAPTOR", "3441", "38.59", "Graph or tree RAG" ], [ "Fast-GraphRAG", "4204", "48.55", "Graph or tree RAG" ], [ "HippoRAG", "7208", "38.52", "Graph or tree RAG" ], [ "LightRAG", "100832", "49.07", "Graph or tree RAG" ] ] }, { "key": "which-customers-are-affected-by-recall-r-17", "title": "Which customers are affected by recall R-17?", "type": "knowledgeGraph", "source": "", "caption": "Illustration. The recall notice names a part, the bill of materials links the part to products, the asset register links products to customers. Chunk retrieval finds the notice; the graph finds the customers.", "csv": "https://bardak.ai/data/graphrag-vs-rag-evidence/which-customers-are-affected-by-recall-r-17.csv", "header": [ "source", "target", "relation" ], "rows": [ [ "Recall R-17", "Capacitor C12", "concerns" ], [ "Supplier Kestrel", "Capacitor C12", "supplies" ], [ "Capacitor C12", "Infusion pump P-200", "used in" ], [ "Capacitor C12", "Monitor M-40", "used in" ], [ "Infusion pump P-200", "Hospital North", "installed at" ], [ "Infusion pump P-200", "Clinic East", "installed at" ], [ "Monitor M-40", "Hospital North", "installed at" ], [ "Monitor M-40", "Care home West", "installed at" ] ] } ] }