ModelRefs / GraphRAG — AI Glossary

GraphRAG — AI Glossary

A RAG variant that retrieves over a knowledge graph — enabling global summary queries that pure vector search cannot answer.

Overview

GraphRAG (Microsoft Research, 2024) clusters documents into entity communities and builds a hierarchical graph. Queries can then use community summaries to answer holistic questions like 'What are the main themes in this corpus?' Vector RAG handles local factual retrieval better.

Reference details

Topic	rag
Last reviewed	2026-06-24

How it works

An indexing pass extracts entities and relationships from the corpus with a model, producing a graph rather than only chunk embeddings.
Related entities are grouped into communities, and a summary is generated per community — so the index holds synthesised overviews as well as source text.
A local query traverses from matched entities along their relationships, which is how a fact two hops away becomes reachable without a second retrieval round.
A global query answers over the community summaries instead of the chunks, which is what makes corpus-wide questions answerable at all.
Indexing is therefore expensive and front-loaded: the model is run over the whole corpus once to build the graph, rather than only embedding it.

Example: Where the indexing cost goes, and what it buys

Embedding 10,000 chunks is 10,000 cheap embedding calls. Building a graph means running a capable model over all 10,000 to extract entities and relations, then generating community summaries — tens of thousands of generation calls, orders of magnitude more expensive. In return, 'what are the main themes across this corpus?' becomes answerable, which no top-k retrieval can address because the answer is in no single chunk. The cost is real and the capability is genuinely different.

Commonly confused with

Retrieves over an entity graph rather than a flat set of chunks, which lets it answer questions no top-k similarity search can: global questions about themes across a whole corpus, and multi-hop questions connecting entities that never co-occur in one chunk. The cost is an extraction and graph-construction step over the corpus up front, so it is not a drop-in change to an existing pipeline.

When to use it

Reach for it when:

Questions are corpus-wide — themes, patterns, summaries across many documents.
Answers require traversing relationships between entities rather than matching passages.
The corpus is stable enough that expensive indexing is amortised.

Reach for something else when:

Questions are single-fact lookups, where vector retrieval answers at a fraction of the cost.
The corpus changes constantly, making repeated graph construction prohibitive.
Indexing budget is the constraint, since the cost is front-loaded and large.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to GraphRAG — AI Glossary.

Frequently asked questions

What is GraphRAG?