# What is Graph RAG?

Also called graph-based retrieval augmented generation.

Graph RAG is retrieval augmented generation that retrieves over a knowledge graph built from the source documents, rather than over independent text chunks alone. Entities and their relationships are extracted during ingestion, and queries traverse those connections to gather evidence spread across many documents. It targets questions that no single passage answers.

Standard chunk retrieval assumes the answer lives inside one or a few passages. That assumption breaks for aggregate and connective questions, such as which suppliers appear across several incident reports, or how two projects relate through shared people. Answering those requires assembling facts that were never written down together, which is what traversing a graph of extracted entities and relations makes possible.

Ingestion is the heavy part. A model reads each document and emits entities, relationships, and supporting quotations, which are merged into a graph where the same real-world entity mentioned differently must be resolved into one node. Many implementations then cluster the graph into communities and pre-generate summaries per community, so broad questions can be answered from summaries instead of raw traversal.

Query time usually blends approaches. A local mode starts from entities mentioned in the question and walks a short distance to collect connected facts; a global mode reasons over community summaries for corpus-wide questions. Text chunks stay attached to graph elements so answers can still cite specific source passages rather than only asserting extracted relations.

The trade-offs are substantial. Ingestion requires many model calls and therefore real cost and time, extraction errors and bad entity resolution propagate into every answer, and the graph must be maintained as documents change. Microsoft's open GraphRAG work in 2024 popularized the approach, but for corpora where answers do sit in single passages, conventional retrieval remains simpler and usually sufficient.

## Key points

- Retrieves over extracted entities and relationships, not only chunks
- Answers connective and corpus-wide questions single passages cannot
- Ingestion cost is high and dominated by extraction model calls
- Entity resolution errors propagate into every answer
- Overkill when answers already live in individual passages

## In practice

A question asks which vendors were involved in more than one outage last year. No single report contains that list. Graph RAG traverses from the outage nodes to their linked vendor nodes, finds two vendors connected to three incidents each, and cites the six underlying reports. Chunk retrieval would have returned a handful of individual incident passages and left the counting to guesswork.

## Related terms

- [Knowledge Graph](/en/glossary/knowledge-graph)
- [Retrieval Augmented Generation](/en/glossary/retrieval-augmented-generation)
- [Entity Resolution](/en/glossary/entity-resolution)
- [Fact Extraction](/en/glossary/fact-extraction)
- [Agentic Retrieval](/en/glossary/agentic-retrieval)

[Back to the AI Glossary](/en/glossary)