
GraphRAG: retrieval over a knowledge graph of your data
In short: GraphRAG builds a knowledge graph of entities and relationships from your documents and retrieves over it to answer broad, connected questions.
GraphRAG is retrieval over a knowledge graph of your data: an LLM first extracts entities and their relationships from documents into a graph, then answers questions from that graph and from summaries of its communities. The paper that proposed it, submitted by Edge and co-authors in April 2024, targets questions plain chunk retrieval handles badly, such as "what are the main themes in the dataset?"
GraphRAG in plain words
Normal RAG finds the right pages. GraphRAG first draws a map of who and what is connected to whom across all pages, then answers from the map.
How it works
- An LLM reads the source documents and extracts entities (people, systems, products) and relationships.
- These form a graph, which is grouped into communities of closely related entities.
- The system pregenerates a summary for each community.
- At query time, each relevant community summary produces a partial answer, and the partial answers are summarised into the final response.
On datasets in the 1 million token range, the paper reports substantial improvements over a conventional RAG baseline in the comprehensiveness and diversity of answers.
The cost sits in indexing, because an LLM processes the whole corpus first. Microsoft's GraphRAG repository warns that indexing can be an expensive operation and recommends tuning the prompts for your data.
Example
As an illustration, a question over a year of incident reports: "which systems keep appearing together in outages?" Chunk retrieval finds individual incidents; GraphRAG's graph shows the payment gateway and the identity service linked across many incidents.
Where you'll meet it
- Microsoft's open-source GraphRAG, MIT licensed, and similar libraries
- Graph databases used to store the extracted entities and relationships
- Investigations, research synthesis and compliance analysis
- Enterprise search where relationships matter more than single facts
What it changes for you
For architects, GraphRAG is a tool for global, relationship-heavy questions, and it should not be the default upgrade. For finance, indexing cost scales with corpus size, so pilot on a subset. For data teams, the extracted graph is itself a useful artefact for analytics, beyond question answering.
GraphRAG vs standard RAG
Standard RAG retrieves similar text chunks, best for specific questions. GraphRAG retrieves structured relationships and summaries, best for broad, connective questions.
Common mistake
My view: choosing GraphRAG because it sounds more advanced. If most user questions are specific lookups, hybrid search with reranking is cheaper and simpler, and I would expect it to answer them as well.
Related terms
Retrieval-augmented generation · Agentic RAG · Hybrid search · Embeddings
From the archive: GraphRAG precomputes answers before questions
Choose GraphRAG for broad, connective questions, and keep hybrid search for specific lookups.


