Overview
GraphRAG addresses corpus-wide questions by extracting entities, relationships, and claims, organizing them into a hierarchical graph, and composing community summaries at query time. The series traces this pipeline, its coverage and cost trade-offs, and reported evidence of stronger comprehensiveness and diversity than vector RAG on studied global questions—not a universal accuracy or replacement guarantee.
Learning outcomes
- GraphRAG fits global questions that require organized coverage across a corpus, not merely a relevant passage.
- Its extraction, graph, hierarchy, and report stages make broad evidence reusable but introduce matching and compression losses.
- In the reported studies, global methods were more comprehensive and diverse than vector RAG on generated sensemaking questions.
- Vector RAG remained more direct, and empowerment results were mixed across the reported comparisons.
- Graph-free map-reduce is a meaningful baseline; GraphRAG’s added indexing cost and organization need workload-specific justification.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
When is GraphRAG a better fit than vector RAG?
GraphRAG is a better fit when a question requires themes, relationships, or patterns across a large corpus rather than one localized fact. Vector RAG remains suited to precise, direct retrieval.
What does GraphRAG add before a user asks a question?
Before query time, GraphRAG extracts entities, relationships, and claims, consolidates them into a graph, detects nested communities, and prepares summaries. Those views can be reused for later global questions.
Does GraphRAG always produce more accurate answers?
No. The reported evaluation found broader and more diverse answers on studied global questions, but its LLM-based measures did not independently verify every claim or establish universal accuracy.
How do GraphRAG’s breadth and cost trade off?
GraphRAG trades upfront indexing work and hierarchy choices for more organized corpus-wide coverage. Root summaries reduce query-time context, while lower levels preserve more detail at greater processing and selection cost.
Is a knowledge graph necessary for global summarization?
No. Graph-free map-reduce also improved global coverage relative to vector RAG in the reported comparisons, so GraphRAG’s additional thematic organization must be justified for the target workload.