---
title: GraphRAG
description: 'A family of retrieval-augmented generation methods that builds or uses graph structure to retrieve relationships, communities, paths, or corpus-level themes.'
canonical_url: 'https://darkfactory.dev/glossary/graphrag'
markdown_url: 'https://darkfactory.dev/glossary/graphrag.md'
collection: glossary
date_published: '2026-08-04T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# GraphRAG


## Definition

GraphRAG is a family of retrieval-augmented generation methods that builds or uses graph structure to retrieve relationships, paths, communities, or themes that isolated passage similarity may miss. The graph can be an existing knowledge graph or a projection extracted from the target documents.

Microsoft Research's specific GraphRAG approach was introduced in a 2024 paper on global questions over large document collections. Its indexer extracts entities and relationships, groups the graph into communities, and generates community summaries. A global query produces partial answers from relevant community reports and combines them. The published evaluation found better comprehensiveness and diversity than its conventional RAG baseline for the tested global sensemaking questions.

Current systems use "GraphRAG" more broadly for several graph-assisted retrieval designs. Some retrieve local neighborhoods around matched entities. Others traverse paths, combine vector and graph search, or route between local and global modes. Name the implementation and query mode instead of assuming one standard pipeline.

Microsoft's paper clearly proposes and names its GraphRAG method. It does not establish that Microsoft coined every earlier or later use of graph-based RAG as a general phrase.

## Distinguish it from nearby terms

- **Conventional RAG** often retrieves independent text passages by keyword or vector similarity. GraphRAG adds explicit graph structure to retrieval or summarization.
- A **knowledge graph** is the structured graph artifact. GraphRAG is the retrieval and generation pipeline that uses it.
- A **vector database** can still participate in GraphRAG for entity matching, text-unit retrieval, or hybrid search.
- GraphRAG is not a replacement for source documents. A model-extracted node, edge, or summary can be wrong.

## Operational significance

Graph construction adds extraction calls, storage, tuning, refresh work, and new failure modes. Preserve links from entities, edges, claims, and summaries back to source spans. Evaluate local factual questions, multi-hop relationship questions, and corpus-level questions separately. Compare against a simpler retrieval baseline because a graph earns its cost only when the target question class benefits from the added structure.

## Check your understanding

A question asks for the main concerns appearing across hundreds of reports. Community summaries may help. A question asks for one invoice number in one page. Direct text retrieval is likely cheaper and easier to verify.

## Also called

graph retrieval-augmented generation, graph-based RAG

## Related terms

- [Retrieval-augmented generation (RAG)](https://darkfactory.dev/glossary/retrieval-augmented-generation)
- [Knowledge graph](https://darkfactory.dev/glossary/knowledge-graph)
- [Vector database](https://darkfactory.dev/glossary/vector-database)
- [Semantic search](https://darkfactory.dev/glossary/semantic-search)

## Related factory areas

- [Context, memory, knowledge & skills](https://darkfactory.dev/factory/context-memory-skills)

## Evidence and further reading

- [Microsoft GraphRAG Documentation](https://microsoft.github.io/graphrag/)
- [From Local to Global: A Graph RAG Approach to Query-Focused Summarization](https://arxiv.org/abs/2404.16130)
- [Turing Post: Is Graph Engineering Real?](https://www.turingpost.com/p/is-graph-engineering-real-why-everyone-is-talking-about-it)