--- id: kb-2026-00019 title: GraphRAG schema_type: TechArticle category: ai language: en confidence: high last_verified: '2026-05-25' created_date: '2026-05-22' generation_method: ai_structured ai_models: - claude-opus derived_from_human_seed: true conflict_of_interest: none_declared is_live_document: false data_period: static atomic_facts: - id: fact-ai-001 statement: >- Unlike standard RAG which retrieves flat text chunks via vector similarity, GraphRAG constructs a knowledge graph from raw text, detects communities using graph algorithms, and generates hierarchical summaries at multiple levels. source_title: 'From Local to Global: A Graph RAG Approach to Query-Focused Summarization' source_url: https://arxiv.org/abs/2404.16130 confidence: medium - id: fact-ai-002 statement: This enables both local (entity-level) and global (dataset-level) query understanding. source_title: 'From Local to Global: A Graph RAG Approach to Query-Focused Summarization' source_url: https://arxiv.org/abs/2404.16130 confidence: medium - id: fact-ai-004 statement: Standard RAG works well for localized factoid questions, while GraphRAG is designed for global queries that require synthesizing information across an entire dataset. source_title: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks source_url: https://arxiv.org/abs/2005.11401 confidence: medium completeness: 0.85 known_gaps: - Statistics and data cited are from 2024 and earlier; more recent developments may have become available since publication - Certain sub-topics are covered at a general level; specialized edge cases and nuanced applications may not be fully addressed disputed_statements: - statement: Graph-based RAG approaches trade computational cost for improved multi-hop reasoning; the cost-benefit ratio varies significantly by domain and task complexity primary_sources: - title: 'From Local to Global: A Graph RAG Approach to Query-Focused Summarization' authors: - Edge - Trinh - Cheng - et al. type: academic_paper year: 2024 url: https://arxiv.org/abs/2404.16130 institution: Microsoft Research - title: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks authors: - Lewis - Perez - Piktus - et al. type: academic_paper year: 2020 url: https://arxiv.org/abs/2005.11401 institution: Facebook AI Research secondary_sources: - title: 'From Local to Global: A Graph RAG Approach to Query-Focused Summarization (Microsoft Research)' type: technical_report year: 2024 authors: - Edge, Darren - Trinh, Ha - Cheng, Newman - et al. institution: Microsoft Research url: https://arxiv.org/abs/2404.16130 - title: 'Knowledge Graphs Meet LLMs: A Comprehensive Survey of Knowledge-Enhanced Language Models' type: survey_paper year: 2024 authors: - multiple institution: IEEE TKDE url: https://doi.org/10.1109/TKDE.2024.3361474 - title: 'RAFT: Adapting Language Model to Domain Specific RAG (UC Berkeley)' type: conference_paper year: 2024 authors: - Zhang, Tianjun - Patil, Shishir G. - Lakshmanan, Laks V. S. - et al. institution: UC Berkeley / ICLR url: https://arxiv.org/abs/2403.10131 updated: '2026-05-24' --- ## TL;DR GraphRAG is a structured, hierarchical approach to Retrieval-Augmented Generation developed by Microsoft Research and open-sourced in 2024. Unlike standard RAG, which retrieves flat text chunks via vector similarity, GraphRAG constructs a knowledge graph from raw text, detects communities using graph algorithms, and generates hierarchical summaries at multiple levels. This enables both local entity-level and global dataset-level query understanding. ## Core Explanation Standard RAG retrieves the top-k most semantically similar text chunks and feeds them to an LLM. This works well for factoid questions ("What is X?") but fails on global queries that require synthesizing information across an entire dataset ("What are the main themes in these 1,000 documents?"). GraphRAG solves this by: 1. **Entity and Relationship Extraction**: LLM extracts entities and relationships from source documents 2. **Knowledge Graph Construction**: Entities become nodes, relationships become edges 3. **Community Detection**: Graph algorithms (Leiden community detection) identify clusters of related entities 4. **Hierarchical Summarization**: LLM generates summaries at community, sub-community, and entity levels 5. **Dual Search Modes**: - **Local Search**: Entity-focused queries use the entity's neighborhood in the graph - **Global Search**: Dataset-level queries use community summaries to answer broad thematic questions ## Development Timeline | Date | Milestone | | ------------ | -------------------------------------------------------------------- | | Feb 13, 2024 | First blog post: "Unlocking LLM discovery on narrative private data" | | Jul 2, 2024 | **Open-sourced on GitHub** | | Sep 9, 2024 | Auto-tuning: automatic parameter optimization for new domains | | Oct 31, 2024 | **DRIFT Search**: combines global and local search | | Nov 15, 2024 | Dynamic community selection for improved global search | | Nov 25, 2024 | **LazyGraphRAG**: cost-optimized variant | | Dec 16, 2024 | **GraphRAG 1.0** official release | | Mar 19, 2025 | Claimify: high-quality claim extraction | | Jun 5, 2025 | BenchmarkQED: automated RAG benchmarking | | Aug 5, 2025 | **VeriTrail**: hallucination detection and attribution | | May 2026 | 33,157 GitHub stars, deployed in Microsoft Discovery | ## Comparison: Standard RAG vs. GraphRAG | Dimension | Standard RAG | GraphRAG | | ---------------- | --------------------------------------------- | -------------------------------------------------------------- | | Data structure | Flat vector chunks | Knowledge graph (entities + relations) | | Retrieval method | Semantic similarity (vector search) | Graph traversal + community detection + vector search | | Query types | Local (factoid) | **Both local and global** (dataset-level synthesis) | | Explainability | Retrieved chunks | Graph paths, community summaries, entity relationships | | Information loss | Chunk boundaries may fragment connected ideas | Entity-relationship model preserves document-level connections | | Cost | Lower (single retrieval + generation) | Higher (graph construction + multi-level summarization) | ## Key Components - **LazyGraphRAG** (Nov 2024): Reduces indexing costs by 100x compared to full GraphRAG while maintaining quality — uses lightweight embedding-based graph construction instead of full LLM extraction - **DRIFT Search** (Oct 2024): Hybrid search combining global community summaries with local entity-level retrieval - **VeriTrail** (Aug 2025): Attribution framework that traces each generated claim back to source documents, detecting hallucinations - **Claimify** (Mar 2025): Extracts structured, verifiable claims from LLM outputs for downstream verification ## Production Deployment Microsoft Discovery (built on Azure) integrates GraphRAG and LazyGraphRAG for scientific research — an agentic platform that helps researchers query and synthesize insights across large document corpora. A Solution Accelerator (API experience on Azure) is available for enterprise deployment, though the original accelerator has been archived in favor of the core library. ## Further Reading - [GraphRAG GitHub](https://github.com/microsoft/graphrag): Open-source repository (33K+ stars) - [Microsoft Research Project Page](https://www.microsoft.com/en-us/research/project/graphrag/): Official overview and publications - [GraphRAG Paper](https://www.microsoft.com/en-us/research/publication/from-local-to-global-a-graph-rag-approach-to-query-focused-summarization/): From Local to Global — A Graph RAG Approach