{
  "post": {
    "title": "Five questions before you build a knowledge graph for RAG",
    "url": "https://bardak.ai/posts/knowledge-graph-rag-five-questions",
    "updated": "2026-10-29T15:31:48.022Z"
  },
  "citation": "Five questions before you build a knowledge graph for RAG, Ivan Bardak, bardak.ai, https://bardak.ai/posts/knowledge-graph-rag-five-questions",
  "datasets": [
    {
      "key": "five-questions-before-you-build-a-graph",
      "title": "Five questions before you build a graph",
      "type": "benchmarkTable",
      "source": "",
      "caption": "Answer all five before choosing a framework. One strong \"vector\" answer can outweigh several \"graph\" ones.",
      "csv": "https://bardak.ai/data/knowledge-graph-rag-five-questions/five-questions-before-you-build-a-graph.csv",
      "header": [
        "Question",
        "Points to a graph",
        "Points to vector or hybrid RAG"
      ],
      "rows": [
        [
          "1. How many of your questions need several hops?",
          "A large share of multi-hop, comparison or time-based questions",
          "Mostly lookups of a fact, a policy or a spec"
        ],
        [
          "2. Do people ask about the whole corpus?",
          "Themes, trends and changes across many documents",
          "Answers live in one or two documents"
        ],
        [
          "3. Can entities be extracted reliably?",
          "Clear, consistent entities: parts, products, customers, contracts",
          "Free text where entities are vague or named inconsistently"
        ],
        [
          "4. How fast does the corpus change?",
          "Stable content, or a budget for regular re-indexing",
          "Daily changes and a tight inference budget"
        ],
        [
          "5. Do permissions differ between documents?",
          "One audience sees everything, or the graph can be split per audience",
          "Fine-grained, per-document access rules"
        ]
      ]
    },
    {
      "key": "tokens-processed-to-index-one-book-about-56-000-tokens",
      "title": "Tokens processed to index one book (about 56,000 tokens)",
      "type": "chart",
      "source": "Xiang et al. 2025, GraphRAG-Bench, Appendix G.3, Table 10",
      "caption": "Microsoft's GraphRAG processed about 12 tokens for every token in the book; HippoRAG2 about 6; RAPTOR about 2. A vector index embeds each chunk once, on a much cheaper embedding model.",
      "csv": "https://bardak.ai/data/knowledge-graph-rag-five-questions/tokens-processed-to-index-one-book-about-56-000-tokens.csv",
      "header": [
        "method",
        "Tokens"
      ],
      "rows": [
        [
          "MS-GraphRAG",
          "654673"
        ],
        [
          "LazyGraphRAG",
          "591150"
        ],
        [
          "KET-RAG",
          "517551"
        ],
        [
          "LightRAG",
          "474172"
        ],
        [
          "HippoRAG2",
          "329993"
        ],
        [
          "Fast-GraphRAG",
          "251817"
        ],
        [
          "HippoRAG",
          "177961"
        ],
        [
          "RAPTOR",
          "115541"
        ],
        [
          "KGP",
          "89215"
        ]
      ]
    }
  ]
}