GraphRAG
Definition
GraphRAG replaces (or supplements) flat chunk retrieval with an LLM-generated knowledge graph over the corpus, then retrieves over that graph’s structure rather than over independent passage embeddings.
The Problem It Addresses
Microsoft Research’s framing identifies two failure modes of baseline RAG that no amount of embedding-quality work fixes:
- Connecting disparate information — the answer requires joining facts that live in different documents, and each document alone scores poorly against the query.
- Holistic semantic understanding — questions about themes across a whole collection (“what are the main topics here?”) have no single passage that answers them.
Both are consequences of scoring each passage independently — a limitation of similarity ranking that no better embedding model removes.
Mechanism
- An LLM extracts entities and relationships from the source documents into a knowledge graph.
- Bottom-up clustering organizes the graph hierarchically into semantic clusters.
- Each cluster is pre-summarized, so thematic questions have something to retrieve against.
- Query time traverses graph structure rather than ranking isolated chunks.
Evidence
Microsoft Research evaluated on the VIINA dataset (Violent Incident Information from News Articles) — thousands of Russian and Ukrainian news articles from June 2023 — using GPT-4 Turbo for graph construction and LangChain’s Q&A as the baseline. GraphRAG consistently outperformed baseline RAG on comprehensiveness, human enfranchisement and diversity, while holding similar faithfulness under SelfCheckGPT evaluation.
The graph machine learning came from the graspologic library.
Related Concepts
- RAG — the baseline this modifies
- HippoRAG — a neurobiologically-motivated variant using Personalized PageRank over the graph
- Knowledge Graph Search — the underlying representation
- Long-Context RAG — the competing answer to the same “not enough context” problem
Articles
- GraphRAG - Unlocking LLM discovery on narrative private data — Jonathan Larson, Steven Truitt; the originating Microsoft Research post