What is GraphRAG?

What GraphRAG is

GraphRAG is a form of retrieval-augmented generation (RAG): a system retrieves information from a collection and uses it as context for a language model’s response. In Microsoft’s documented approach, the index includes a graph, a structure that represents entities and their relationships, plus other material such as text chunks, claims, or summaries. The graph is built from source text to help organize and retrieve information. It is not the original collection itself. GraphRAG is also used as the name of Microsoft’s particular open project, so descriptions of its indexing steps refer to that documented implementation.

How it works

The process has an indexing stage and a question-answering stage. During indexing, the documented standard pipeline extracts entities and relationships from text units, groups related entities into communities, and creates summaries or reports for those communities. It can also embed text, meaning it represents text as numerical vectors for similarity search. These prepared outputs form an index that can be searched later; the system does not need to build the graph anew for every question.

When a question arrives, a retrieval method chooses information to include in the language model’s context, the material available to shape its response. Local search identifies relevant entities and may bring in connected entities, relationships, community reports, and related source text. Global search looks across community reports to form an answer about the collection as a whole. The chosen method and the information it retrieves affect what evidence the model sees.

Where it fits among retrieval methods

GraphRAG sits within the broader family of retrieval-augmented generation. A basic RAG system can retrieve text chunks using vector similarity, which compares numerical representations of the question and text. GraphRAG adds graph structure and, in some designs, summaries of groups of related entities. This can suit questions about a particular entity’s connections or broad themes across a collection. Whether it helps depends on the question, the source material, and how the index and retrieval are built.

The methods can coexist. Microsoft’s query documentation describes basic vector RAG alongside local and global search. So “GraphRAG vs RAG” is not always a choice between mutually exclusive systems: GraphRAG is a graph-based approach within the RAG family, and a GraphRAG system may still use vector retrieval. For background on vector search, see What is a vector database?.

Common confusions

GraphRAG does not mean that an answer comes from a graph alone. In Microsoft’s local-search method, the system combines graph information with text chunks from the original documents. The graph can act as a way into connected entities and relationships; source text can supply additional detail. Different GraphRAG designs and search methods may select context differently, so the name does not describe one fixed sequence of steps for every system.

A graph node, relationship, or summary is an extracted or generated part of an index. Its presence shows what the indexing process represented from the source material; it does not independently verify that representation. Likewise, a generated answer is based on the context selected for it. Graph structure can help organize evidence, but the structure alone does not establish that the answer is factually correct.