← AI Terminology
GraphRAG
GraphRAG builds a knowledge graph (entities, relations, communities) from a corpus and retrieves via graph structure plus text — improving global questions over plain chunk RAG.
Microsoft’s GraphRAG popularised the pattern for corpus-level summarisation QA.
Microsoft’s GraphRAG popularised the pattern for corpus-level summarisation QA.
Why It Matters in AI
Chunk RAG struggles with “themes across the whole dataset” questions. Graphs and community summaries provide global context and multi-hop structure. It is heavier to index but powerful for investigative and enterprise corpora.
Key Points
| Aspect | Description |
|---|---|
| Cost | Expensive indexing; LLM extraction at scale |
| Origin | Microsoft GraphRAG project (2024) |
| Related | Knowledge graphs, agentic RAG, multi-hop |
| Pipeline | Extract entities/relations → graph → community summaries → query-time traversal/retrieval |
| Strength | Global themes, multi-hop relations |
| Vs vector RAG | Better global; local factoid QA may not need it |
Simple Analogy
Not only searching loose paragraph cards, but also a wall map of people and companies with strings between them — follow connections for big-picture questions.
Common Usage Examples
- Microsoft GraphRAG open-source pipeline
- Community summary reports for “overview” queries
- Entity graphs in Neo4j + LLM extractors
- Compare global QA vs naive chunk RAG
Summary
In short: GraphRAG retrieves with a knowledge graph and community summaries — stronger global and multi-hop answers than chunk-only RAG.