https://www.youngju.dev/blog 천천히 올바르게. AI Researcher & DevOps Engineer Youngju's tech blog. GPU/CUDA, LLM, MLOps, Kubernetes AI workloads, distributed training, and data engineering. ko fjvbn2003@gmail.com (Youngju Kim) fjvbn2003@gmail.com (Youngju Kim) Wed, 15 Jul 2026 00:00:00 GMT https://www.youngju.dev/blog/2026-07-15-graph-rag-explained.en https://www.youngju.dev/blog/2026-07-15-graph-rag-explained.en The standard RAG recipe — chunk, embed, retrieve top-k — works when the answer sits inside a single chunk, but it stalls structurally on multi-hop questions and on global sensemaking questions that span the whole corpus (what are the main themes?). Microsoft's GraphRAG targets those two blind spots: it uses an LLM to extract a knowledge graph from the corpus, detects communities with the Leiden algorithm, and pregenerates community summaries. Global questions are then answered by map-reduce over those summaries; local questions by entity-centric graph traversal. This post covers where vector RAG breaks, how the GraphRAG pipeline actually works (graph construction, community detection, local vs global search), a concrete example, and the honest tradeoffs — heavy indexing LLM cost, latency, re-indexing pain, the fact that it is often overkill, and that hybrids with vector RAG are common. Not hype — a practical guide to whether you should use it. Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) raggraph-ragknowledge-graphaillm https://www.youngju.dev/blog/2026-07-15-graph-rag-explained.ja https://www.youngju.dev/blog/2026-07-15-graph-rag-explained.ja RAGの標準レシピ(チャンク→埋め込み→上位k件の検索)は、答えが一つのチャンクに収まっているときはよく効きますが、マルチホップの質問や、コーパス全体を貫くグローバルな「センスメイキング」質問では構造的に行き詰まります。MicrosoftのGraphRAGはこの2つの死角を狙い、LLMでコーパスから知識グラフを抽出し、Leidenアルゴリズムでコミュニティを検出して要約を事前生成します。そのうえでグローバルな質問はコミュニティ要約のマップリデュースで、ローカルな質問はエンティティ中心のグラフ探索で答えます。本稿では、ベクトルRAGが行き詰まる場所、GraphRAGパイプラインの実際の動作(グラフ構築→コミュニティ検出→ローカル/グローバル検索)、具体例、そして正直なトレードオフ(索引のLLMコスト、レイテンシ、再索引の負担、しばしば過剰であること、ベクトルとのハイブリッドが一般的であること)を整理します。ハイプではなく、あなたがこれを使うべきかに答える実務ガイドです。 Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) raggraph-ragknowledge-graphaillm https://www.youngju.dev/blog/2026-07-15-graph-rag-explained https://www.youngju.dev/blog/2026-07-15-graph-rag-explained RAG의 표준 레시피(청크 → 임베딩 → 상위 k개 검색)는 답이 한 조각에 들어 있을 때 잘 작동하지만, 멀티홉 질문과 코퍼스 전체를 관통하는 글로벌 센스메이킹 질문에서는 구조적으로 막힙니다. Microsoft의 GraphRAG는 이 두 사각지대를 겨냥해, LLM으로 코퍼스에서 지식 그래프를 뽑고 Leiden 알고리즘으로 커뮤니티를 찾아 요약을 미리 만들어 둡니다. 그런 다음 글로벌 질문은 커뮤니티 요약의 맵-리듀스로, 로컬 질문은 엔티티 중심 그래프 탐색으로 답합니다. 이 글은 벡터 RAG가 막히는 지점, GraphRAG 파이프라인의 실제 동작(그래프 구축 → 커뮤니티 탐지 → 로컬 vs 글로벌 검색), 구체적인 예, 그리고 정직한 트레이드오프(색인 LLM 비용, 지연, 갱신 부담, 자주 과잉이라는 점, 벡터와의 하이브리드가 흔하다는 점)를 정리합니다. 하이프가 아니라 당신이 이걸 써야 하는지에 답하는 실무 가이드입니다. Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) raggraph-ragknowledge-graphaillm https://www.youngju.dev/blog/2026-07-15-graph-rag-explained.zh https://www.youngju.dev/blog/2026-07-15-graph-rag-explained.zh RAG 的标准配方(分块 → 嵌入 → 检索前 k 个)在答案完整落在某一个片段里时很好用，但在多跳问题、以及贯穿整个语料库的全局「意义建构」问题上会结构性地卡住。Microsoft 的 GraphRAG 正是瞄准这两处盲区 — 用 LLM 从语料库里抽取知识图谱，用 Leiden 算法找出社区并预先生成摘要。此后，全局问题用社区摘要的映射-归约来回答，局部问题则用以实体为中心的图探索来回答。本文梳理了向量 RAG 卡住的地方、GraphRAG 流水线的实际运作(构建图 → 社区检测 → 局部 vs 全局检索)、具体的例子，以及诚实的权衡(索引阶段的 LLM 成本、延迟、重建索引的负担、它经常是过度的、以及与向量的混合方案很常见这几点)。这不是炒作，而是一份回答「你到底该不该用它」的实务指南。 Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) raggraph-ragknowledge-graphaillm https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks.en https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks.en Graph databases, the RDF and ontology stack, and graph RAG frameworks have exploded over the last two years. Link lists are already abundant, so this post draws an honestly-opinionated map instead. For each category: what it is for, two or three representative tools, and a clear when-to-pick-what — embedded (Kùzu) vs server (Neo4j), formal OWL/RDF vs pragmatic property graph, a roll-your-own extraction pipeline vs Microsoft GraphRAG's batteries-included one. And the most important question: when NOT to use a graph at all, because plain vector RAG is enough. It closes the series with a small decision flow. Every tool here was verified to exist, and its maturity and positioning are represented honestly — including the fact that Kùzu was archived in 2025. Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) toolsknowledge-graphgraph-ragontologyframeworks https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks.ja https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks.ja グラフデータベース、RDF・オントロジースタック、グラフRAGフレームワークは、この2年で爆発的に増えました。リンク集はもう十分にあるので、この記事は代わりに、正直に意見を込めた地図を描きます。カテゴリごとに、何のためのものか、代表的なツールを2〜3個、そして「いつ何を選ぶか」をはっきり述べます — 組み込み(Kùzu)vsサーバー(Neo4j)、形式的なOWL/RDF vs実用的なプロパティグラフ、自作の抽出パイプライン vs Microsoft GraphRAGのバッテリー同梱パイプライン。そして最も重要な問い: そもそもグラフを使うべきでないとき(ベクトルRAGで十分なとき)はいつか。小さな決定フローでシリーズを締めます。すべてのツールは実在を確認し、成熟度とポジショニングを誇張していません — Kùzuが2025年にアーカイブされた事実も含めて。 Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) toolsknowledge-graphgraph-ragontologyframeworks https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks 그래프 데이터베이스, RDF·온톨로지 스택, 그래프 RAG 프레임워크는 지난 2년 사이 폭발적으로 늘었습니다. 링크 목록은 이미 많으니, 이 글은 대신 정직하게 의견을 담은 지도를 그립니다. 카테고리마다 무엇에 쓰는지, 대표 도구 두세 개, 그리고 "언제 무엇을 고를지"를 분명히 말합니다 — 임베디드(Kùzu) vs 서버(Neo4j), 형식 OWL/RDF vs 실용 프로퍼티 그래프, 직접 짜는 추출 파이프라인 vs Microsoft GraphRAG의 배터리 포함 파이프라인. 그리고 가장 중요한 질문: 애초에 그래프를 쓰지 말아야 할 때(벡터 RAG로 충분한 경우)는 언제인가. 작은 결정 흐름으로 시리즈를 닫습니다. 모든 도구는 실제로 존재하는지 확인했고, 성숙도와 포지셔닝을 부풀리지 않았습니다 — Kùzu가 2025년 아카이브된 사실까지 포함해서. Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) toolsknowledge-graphgraph-ragontologyframeworks https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks.zh https://www.youngju.dev/blog/2026-07-15-knowledge-graph-tools-frameworks.zh 图数据库、RDF 与本体栈、图 RAG 框架，在过去两年里爆发式增长。链接清单已经太多，所以本文换一种做法，画一张带着诚实观点的地图。每个类别都讲清楚：拿它做什么、两三个代表性工具，以及「什么时候该选什么」— 嵌入式(Kùzu) vs 服务器(Neo4j)、形式化的 OWL/RDF vs 实用的属性图、自己搭的抽取流水线 vs Microsoft GraphRAG 那种自带电池的流水线。以及最重要的一个问题：什么时候压根就不该用图(向量 RAG 就够了的时候)。本文以一张小小的决策流程为整个系列收尾。所有工具都核实过确实存在，也没有夸大它们的成熟度与定位 — 包括 Kùzu 在 2025 年被归档这一事实。 Wed, 15 Jul 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) toolsknowledge-graphgraph-ragontologyframeworks