{
  "format": "agent-article-v0",
  "article_id": "EA-069",
  "canonical_url": "/tools/ea-069-knowledge-graph-llm",
  "page_type": "reference",
  "claim_ids": [],
  "evidence_ids": [
    "EVD-EA-069-URL-1B0E5BB5CDA6",
    "EVD-EA-069-URL-3643309247ED",
    "EVD-EA-069-URL-50AF901204D5",
    "EVD-EA-069-URL-01640D72AF35",
    "EA-069-R11-RESEARCH",
    "EA-069-R11-WEB-READBACK"
  ],
  "unknown_ids": [
    "EA-069-UNKNOWN-01",
    "EA-069-UNKNOWN-CURRENTNESS"
  ],
  "operational_object_ids": [],
  "provider": [
    "Microsoft"
  ],
  "models": [],
  "plans": [],
  "surfaces": [],
  "regions": [],
  "current_status": "UNKNOWN",
  "observation_status": "SPEC_ONLY",
  "freshness": "UNKNOWN",
  "last_verified": null,
  "supersedes": [],
  "superseded_by": [],
  "provenance": [
    "COMM-065-EA0069",
    "SRC-EA-069-URL-1B0E5BB5CDA6",
    "SRC-EA-069-URL-3643309247ED",
    "SRC-EA-069-URL-50AF901204D5",
    "SRC-EA-069-URL-01640D72AF35",
    "WEBBODY-EA0069-R8-0A978A50"
  ],
  "article_sources": [
    {
      "id": "EVD-EA-069-URL-1B0E5BB5CDA6",
      "evidence_class": "RESEARCH",
      "title": "author/org: Darren Edge et al. / Microsoft-affiliated authors",
      "href": "https://arxiv.org/abs/2404.16130",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-069-URL-1B0E5BB5CDA6"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "canonical URL: https://arxiv.org/abs/2404.16130"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-069-URL-3643309247ED",
      "evidence_class": "RESEARCH",
      "title": "author/org: Microsoft Research",
      "href": "https://www.microsoft.com/en-us/research/project/graphrag/",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-069-URL-3643309247ED"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "canonical URL: https://www.microsoft.com/en-us/research/project/graphrag/"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-069-URL-50AF901204D5",
      "evidence_class": "RESEARCH",
      "title": "author/org: Amazon Web Services",
      "href": "https://aws.amazon.com/about-aws/whats-new/2025/03/amazon-bedrock-knowledge-bases-graphrag-generally-available/",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-069-URL-50AF901204D5"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "canonical URL: https://aws.amazon.com/about-aws/whats-new/2025/03/amazon-bedrock-knowledge-bases-graphrag-generally-available/"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-069-URL-01640D72AF35",
      "evidence_class": "RESEARCH",
      "title": "author/org: Neo4j",
      "href": "https://neo4j.com/docs/neo4j-graphrag-python/current/",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-069-URL-01640D72AF35"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "canonical URL: https://neo4j.com/docs/neo4j-graphrag-python/current/"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EA-069-R11-RESEARCH",
      "evidence_class": "RESEARCH",
      "title": "EA-069 監査済みresearch / canonical revision binding",
      "supports": [
        "- GraphRAGは単一方式ではない。少なくとも「LLMで非構造文書からgraphを派生しcommunity/global summaryへ使う方式」と「既存/生成graphをvector retrieval + graph traversalへ使う方式」を分けるべきである。\n- Microsoft primary researchが直接確立するのは、特定のglobal sensemaking query classでのnaive RAG比改善であり、一般QA・transactional lookup・全domainでの優越ではない。\n- AWSのGA実装は、vectorとgraph traversalのhybrid retrievalが商用managed pathとして存在することを確立する。ただしregion/service dependencyがavailability gateになる。\n- Knowledge graphはground truthではない。LLM-derived entity/relation/claimはsourceから派生したindex artifactであり、原典locatorと抽出model/prompt/versionをprovenanceとして保持すべきである。\n- authoritative KG（人手/業務systemで管理されたentity/relation）とLLM-derived KGはorigin・信頼度・更新責任を分離する必要がある。"
      ],
      "limit": "canonical監査・revision bindingの存在を示す。外部ソースcurrentnessが本文またはcontrol-planeで確定していない場合は、currentnessをUNKNOWNのまま保持する。",
      "origin_class": "AUDITED_CANONICAL_RESEARCH_PACKET",
      "source_or_experiment_ids": [
        "COMM-065-EA0069"
      ],
      "supporting_locators": [
        "COMM-065-EA0069",
        "EA-069-R5-LEGACY-MIGRATION-20261008-R1",
        "content/staging/EA-069.md"
      ],
      "scope": "監査済みEA本文・canonical revision・既存Source Graph relationの範囲。欠損currentnessは推測で閉じない。",
      "provenance_status": "AUDIT_PASS_CANONICAL_BINDING"
    },
    {
      "id": "EA-069-R11-WEB-READBACK",
      "evidence_class": "IMPL",
      "title": "EA-069 Web body identity / canonical readback",
      "supports": [
        "canonical staging bodyのSHA-256一致とR9 canonical ledger readbackをProduction projectionの入力として検証する。"
      ],
      "limit": "body identity/readbackは本文claim自体の真偽や外部ソースの最新性を追加保証しない。",
      "checked_at": null,
      "origin_class": "IMPLEMENTATION_READBACK",
      "source_or_experiment_ids": [
        "WEBBODY-EA0069-R8-0A978A50"
      ],
      "supporting_locators": [
        "WEBBODY-EA0069-R8-0A978A50",
        "content/staging/EA-069.md",
        "EA-069-R5-LEGACY-MIGRATION-20261008-R1"
      ],
      "scope": "R9/R10 canonical readback後のProduction生成時body identity。",
      "provenance_status": "CANONICAL_READBACK_BODY_HASH_MATCH"
    }
  ],
  "research_sections": [
    {
      "heading": "WHY_THIS_MATTERS_TO_UPDATE",
      "text": "Vector similarityだけでは、局所的に近い断片の検索はできても、文書集合全体の関係、複数hopの依存、entity間接続、global sensemakingを直接表現しにくい。GraphRAG/KG-RAGはこれを補完し得るが、LLM抽出graphは原典そのものではなく派生表現であり、誤抽出・entity resolution・staleness・ACL・index構築費用を新しいfailure surfaceとして持つ。将来版では「graphを使う=正確になる」とせず、どのquery classに何を追加する技術かを限定して扱う必要がある。"
    },
    {
      "heading": "SCOPE_ROUTING",
      "text": "Enterprise AI Core / implementation / retrieval & knowledge architecture / benchmark-evaluation / governance / provider mapping / interoperability"
    },
    {
      "heading": "CLAIM_OR_RESEARCH_QUESTION",
      "text": "1. Knowledge graph/GraphRAGは、vector-only RAGが弱いglobal/corpus-level・relationship-heavy queryを補完できるか。\n2. LLM-derived graphとauthoritative enterprise graphを同一視せず、provenance・freshness・ACL・schema/ontologyをどう保持すべきか。\n3. graph retrievalの改善claimを、特定benchmark/query classから一般的な製品保証へ外挿せず評価するには何を測るべきか。"
    },
    {
      "heading": "EVIDENCE TABLE",
      "text": "1) source class: primary paper\nexact title: From Local to Global: A Graph RAG Approach to Query-Focused Summarization\nauthor/org: Darren Edge et al. / Microsoft-affiliated authors\ncanonical URL: https://arxiv.org/abs/2404.16130\npublication/update/access date: v1 2024-04-24; v2 2025-02-19; accessed 2026-09-15\nversion/release/tag/commit/API revision/model ID: arXiv:2404.16130v2\nsupporting locator: abstract / method / evaluation\nlabel: RESEARCH\nrelation: support + limit\nnote: LLM-derived entity graph + community summariesを用い、約1M-token規模の対象でglobal sensemaking questionに対しnaive RAG baselineよりcomprehensiveness/diversity改善を報告。評価対象は限定されたquery class/datasetsであり、全RAG workloadへの一般化は不可。\n\n\n2) source class: original research/project documentation\nexact title: Project GraphRAG\nauthor/org: Microsoft Research\ncanonical URL: https://www.microsoft.com/en-us/research/project/graphrag/\npublication/update/access date: accessed 2026-09-15\nversion/release/tag/commit/API revision/model ID: project current page; exact software release not established in this packet\nsupporting locator: project overview / GraphRAG research timeline\nlabel: SOURCE_REPORT / OSS_PRACTICE\nrelation: context + implementation\nnote: text extraction、network analysis、LLM prompting/summarizationを統合するend-to-end systemとしてGraphRAGを位置付ける。research project/OSS methodologyの存在と、特定enterprise deploymentのSLA/supportを混同しない。\n\n\n3) source class: provider official product announcement\nexact title: Amazon Bedrock Knowledge Bases supports GraphRAG now generally available\nauthor/org: Amazon Web Services\ncanonical URL: https://aws.amazon.com/about-aws/whats-new/2025/03/amazon-bedrock-knowledge-bases-graphrag-generally-available/\npublication/update/access date: 2025-03-07; accessed 2026-09-15\nversion/release/tag/commit/API revision/model ID: Bedrock Knowledge Bases GraphRAG GA; Neptune Analytics dependency\nsupporting locator: announcement paragraphs describing embeddings + entity/relationship graph + vector similarity + graph traversal; region gate\nlabel: VERIFIED_SPEC / PRODUCT_VOLATILE\nrelation: support + availability limit\nnote: GraphRAGはAmazon Neptune Analyticsにvector embeddingsとentity/relationship graphを保存し、vector similarityとgraph traversalを組み合わせる。Bedrock Knowledge BasesとNeptune Analytics双方が利用可能なregionに限定。\n\n\n4) source class: author-maintained implementation documentation\nexact title: GraphRAG for Python\n author/org: Neo4j\ncanonical URL: https://neo4j.com/docs/neo4j-graphrag-python/current/\npublication/update/access date: accessed 2026-09-15\nversion/release/tag/commit/API revision/model ID: current docs; exact package version not fixed in this packet\nsupporting locator: current GraphRAG Python documentation\nlabel: OSS_PRACTICE / PRODUCT_VOLATILE\nrelation: implementation context\nnote: graph databaseを使うGraphRAG実装経路が現行で提供されることを確認。特定version behaviorは後続Workでpackage release/tagを再固定する。"
    },
    {
      "heading": "FINDINGS",
      "text": "- GraphRAGは単一方式ではない。少なくとも「LLMで非構造文書からgraphを派生しcommunity/global summaryへ使う方式」と「既存/生成graphをvector retrieval + graph traversalへ使う方式」を分けるべきである。\n- Microsoft primary researchが直接確立するのは、特定のglobal sensemaking query classでのnaive RAG比改善であり、一般QA・transactional lookup・全domainでの優越ではない。\n- AWSのGA実装は、vectorとgraph traversalのhybrid retrievalが商用managed pathとして存在することを確立する。ただしregion/service dependencyがavailability gateになる。\n- Knowledge graphはground truthではない。LLM-derived entity/relation/claimはsourceから派生したindex artifactであり、原典locatorと抽出model/prompt/versionをprovenanceとして保持すべきである。\n- authoritative KG（人手/業務systemで管理されたentity/relation）とLLM-derived KGはorigin・信頼度・更新責任を分離する必要がある。"
    },
    {
      "heading": "LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE",
      "text": "- graph extraction error: entity/relationship/claimの誤抽出、欠落、重複、entity resolution誤り。\n- stale graph: source更新後もnode/edge/community summaryが古いまま残る。EA-067 Knowledge Freshness Engineeringへ接続。\n- authorization drift: source ACLとgraph node/edge/derived summaryのpermission propagationが一致しない可能性。EA-066 ACL-aware RAGへ接続。\n- derived-summary contamination: community summaryは原典ではなく生成派生物。summaryが誤っていてもgraph traversal自体は成功し得る。\n- ontology/schema mismatch: relation typeの意味がdomain ontologyと一致しない場合、構造化したことで誤りが固定化され得る。\n- cost/latency: graph construction・entity extraction・community detection・summary生成にupfront costがある。Microsoft Research自身もindex構築費とのtrade-offを明示。\n- benchmark transfer: global sensemaking改善をpoint lookup、SQL correctness、regulatory QA、tenant-specific production accuracyへ外挿しない。\n- graph path存在は因果関係の証明ではない。「relationship」「dependency」「causality」は別semantic classとして扱う。"
    },
    {
      "heading": "AVAILABILITY_GATES",
      "text": "AWS managed GraphRAGはBedrock Knowledge BasesとNeptune Analytics双方のregion availabilityに依存する。Neo4j経路はdeployment/package/version/LLM provider構成に依存。Microsoft Research GraphRAGはresearch/OSS methodologyとして扱い、enterprise product entitlement/supportとは別claimにする。PRODUCT_VOLATILE項目は将来更新時にregion・release・support statusを再確認する。"
    },
    {
      "heading": "IMPLEMENTATION_STATE",
      "text": "- Microsoft GraphRAG: research + OSS methodologyとして実装可能。\n- AWS Bedrock Knowledge Bases GraphRAG: 2025-03-07 GA announcementを確認。Neptune Analytics利用。\n- Neo4j GraphRAG Python: current implementation docs存在を確認。\n- provider横断で同一graph schema/query semantics/provenance contractが保証される証拠は未確立。"
    },
    {
      "heading": "BENCHMARK_OR_EVAL",
      "text": "Future evaluationはvector-only vs graph-only vs hybridを同一corpus・同一model条件で比較し、query classを最低でも point/local / multi-hop relationship / global sensemaking / temporal-freshness / unauthorized-access に分ける。metrics: retrieval recall/precision、path correctness、answer correctness、citation/source coverage、unsupported edge rate、entity-resolution error、stale-edge rate、ACL false-allow、latency、indexing/query cost。LLM-as-judgeだけで閉じず、gold relation/pathを持つsynthetic fixtureを併用する。"
    },
    {
      "heading": "SECURITY/PRIVACY",
      "text": "graphは複数文書を横断して新しい関係を顕在化させ得るため、個々のsourceを読めることとcross-source inferenceを許可することを同一視しない。ACL-aware node/edge/summary filteringが必要。\nAUDIT: answer→retrieved node/edge→source chunk/documentまで逆引きできるprovenanceを保持する。"
    },
    {
      "heading": "COST",
      "text": "upfront graph extraction/index/community summarization費とquery-time traversal/LLM費を分離計測する。"
    },
    {
      "heading": "OPERABILITY",
      "text": "source change時のincremental graph update、delete propagation、entity merge/split、schema migration、rebuild条件を運用契約化する。"
    },
    {
      "heading": "SAFE_WORDING_FOR_FUTURE_BOOK",
      "text": "「Knowledge Graph/GraphRAGは、文書間のentity・relationshipを構造化し、vector similarityだけでは扱いにくいmulti-hopやcorpus-levelの探索を補完できる。ただしgraphがLLMから生成される場合、そのnode/edge/summaryは原典から派生した推論・索引artifactであり、正しさは自動保証されない。導入時はprovenance、ACL、freshness、schema/ontology、抽出誤り、index構築費を独立に管理し、query class別にvector-only/hybridとの差を評価する。」"
    },
    {
      "heading": "UPDATE_WORK_HANDOFF",
      "text": "Core候補として `GRAPH KNOWLEDGE CONTRACT` を検討する。最低field: graph_origin(authoritative/derived)、source_id+locator、entity_id/type、edge_type、edge_semantics(association/dependency/causal-claim等)、confidence、extraction_model/prompt/version、created_at/source_version、ACL principals/policy reference、valid_from/to、superseded_by、citation/provenance、delete/tombstone state。\nTEST_PROFILE（未実施）: synthetic corpusにA-B-C multi-hop、cross-document dependency、明示的non-edge、conflicting relation、source update/delete、ACL-separated documentを埋め込み、vector-only / graph-only / hybridで比較。false edge、missing edge、wrong entity merge、stale relation、unauthorized path、answer/citation driftを測る。"
    },
    {
      "heading": "OPEN_UNKNOWN",
      "text": "(1) provider横断でportableなGraphRAG schema/provenance contract、(2) enterprise authoritative KGとLLM-derived KGを統合する際の標準的confidence/override semantics、(3) current provider別ACL propagationとdelete/update SLA、(4) Neo4j等の正確な現行release/tag/commit、(5) graph retrievalがproduction workloadでvector-onlyを上回る境界条件。"
    },
    {
      "heading": "STOP_CONDITION",
      "text": "EA-069については、primary research + current managed product + implementation practice + failure envelope + future TEST_PROFILEまで揃ったためbounded packetとして停止。将来更新WorkでPRODUCT_VOLATILEとversion/regionを再確認する。\n\n---"
    }
  ],
  "human_url": "/tools/ea-069-knowledge-graph-llm",
  "markdown_url": "/agent/article.md?path=%2Ftools%2Fea-069-knowledge-graph-llm",
  "json_url": "/agent/article.json?path=%2Ftools%2Fea-069-knowledge-graph-llm"
}