THE SHORT ANSWER
Standard Vector RAG is fundamentally built for **Local Retrieval**: finding specific needle-in-a-haystack passages that are semantically close to a specific query (e.g. 'What is the refund window for Product X?'). However, when asked **Global Holistic Questions** (e.g. 'What are the top 5 macroeconomic risks across all 500 company earnings transcripts?' or 'What are the main corruption patterns in the Panama Papers?'), vector search completely fails: there is no single chunk that matches the query, and fetching top-50 vector matches only captures a tiny, biased fraction of the dataset. Microsoft Research created **GraphRAG**: (1) An LLM extracts entities and relationships from raw documents into a unified Knowledge Graph, (2) The **Leiden Graph Clustering Algorithm** partitions the graph into hierarchical semantic communities, and (3) An LLM generates pre-computed **Community Summaries** at each hierarchical level. Global queries execute parallel map-reduce syntheses across community summaries, enabling profound dataset-wide sensemaking with 100% comprehensive coverage.
Engineering Handbook & Failure Dynamics
1. Underlying Mechanism
GraphRAG indexation and query pipeline operates across five coordinated stages: (1) Entity & Claim Extraction: LLM parses document chunks, extracting typed entities (Persons, Organizations, Tech) and directed relationships with claim descriptions. (2) Graph Construction: Resolves entity synonyms and builds a unified knowledge graph. (3) Leiden Community Detection: Hierarchical graph clustering (Leiden algorithm) detects modular communities at multiple granularities (Level 0 fine clusters to Level 2 macro themes). (4) Community Report Generation: An LLM summarizes the key narratives, key actors, and structural significance of every community into a standalone report. (5) Global Map-Reduce Query: When a global question is asked, community reports are scored in parallel, intermediate answers are generated in map tasks, and a final executive summary is reduced into a pristine answer.
2. Appropriate Use Context
Investigative journalism on massive document dumps, enterprise compliance auditing across thousands of contracts, biopharmaceutical research literature discovery, and intelligence analysis.
3. Production Failure Modes
Indexation Cost Shock: running entity extraction and community summarization with unoptimized expensive models (GPT-4) on 100,000 documents, costing $40,000+ in indexing fees; community reports containing stale information when source documents are updated without incremental graph synchronization.
4. Diagnostic Signals & Telemetry
Vector RAG answers returning superficial single-document anecdotes when asked high-level summary questions; users rating dataset-wide synthesis queries as 1-star; GraphRAG indexing pipelines completing successfully with high modularity scores in Leiden clustering.
5. Prevention & Safeguards
Use fast, cost-effective models (e.g. GPT-4o-mini, Claude 3.5 Haiku) for bulk entity extraction and community summarization; implement dynamic routing: route specific lookup queries to Local Vector/Graph search and dataset-wide questions to Global Community Map-Reduce; use incremental graph partitioning for live document ingestion.
6. Architectural Trade-offs
GraphRAG requires substantial upfront indexation time and LLM extraction compute, but solves the fundamental blindspot of vector search: comprehensive, corpus-wide global sensemaking.
Case Study (TinyCTO In-Field Example)
A financial intelligence firm analyzed 4,000 leaked corporate audit emails. When analysts asked: 'What are the main systemic compliance violations discussed across all subsidiaries?', traditional vector RAG returned 5 random email snippets discussing a single travel expense discrepancy. The firm deployed Microsoft GraphRAG: the Leiden algorithm detected 18 distinct corporate network communities, generating hierarchical reports. The Global Query mapped across all 18 reports and synthesized a complete 4-point breakdown of cross-border tax avoidance schemes that spanned 6 countries, uncovering insights invisible to vector search.
Interactive Concept Drills
2 CardsWhy does standard Vector RAG fail on global dataset-wide questions?
How does GraphRAG use the Leiden algorithm to organize knowledge?
GraphRAG: Knowledge Graph Extraction, Leiden Community Summaries & Global Sensemaking — Technical FAQ
What is the difference between GraphRAG 'Local Search' and 'Global Search'?
Local Search traverses entity neighborhoods for specific entity questions (e.g. 'Who is Person X?'); Global Search executes map-reduce over community summaries for dataset-wide themes (e.g. 'What are the main topics?').
How can developers reduce GraphRAG indexation costs?
By using fast, cheap extraction models (like GPT-4o-mini or Claude 3.5 Haiku) for entity extraction and prompt-tuning the chunk size to 600-800 tokens to maximize entity density.
🤖 AEO & Key Facts Summary
Key Architectural Facts
- ▸Vector RAG excels at local lookup but fails at global dataset-wide summarization.
- ▸GraphRAG builds an entity knowledge graph and clusters it via the Leiden algorithm.
- ▸Generates pre-computed hierarchical community summaries across the entire corpus.
- ▸Global Map-Reduce queries synthesize comprehensive dataset-wide insights with zero blindspots.
Common Misconceptions
- ✗Misconception: GraphRAG completely replaces vector search (False: GraphRAG combines local vector/graph search with global community summaries for hybrid mastery).
- ✗Misconception: GraphRAG requires manual schema design (False: The LLM dynamically extracts entities and relations directly from unstructured text).
Decision & Governance Guidance
Deploy GraphRAG for document collections requiring high-level synthesis, auditing, and investigative search. Use dynamic query classification: route entity questions to Local Search and thematic questions to Global Search.
Authoritative Sources & Standards
- [OFFICIAL_DOCUMENTATION]From Local to Global: A Graph RAG Approach to Query-Focused Summarization— Darren Edge et al. (Microsoft Research / arXiv)