What is GraphRAG?
GraphRAG is a retrieval-augmented generation approach from Microsoft Research that uses a language model to build a knowledge graph of entities and relationships from a corpus, then answers questions from the graph and pre-written community summaries.
Checked
Vector RAG finds passages similar to the question. That works for "what is the refund window?" and fails for "what are the main themes across these 2,000 interviews?", because no single passage holds that answer.
Microsoft Research introduced GraphRAG in a blog post on 2024-02-13 and described it fully in From Local to Global: A Graph RAG Approach to Query-Focused Summarization by Darren Edge and colleagues, posted to arXiv on 2024-04-24. On global questions over datasets around 1M tokens, the authors report clear gains over a conventional RAG baseline in the comprehensiveness and diversity of answers.
How GraphRAG builds its index
Extract
A language model reads every chunk and pulls out entities, relationships and claims, which form a knowledge graph.
Cluster
The graph is partitioned into a hierarchy of communities of closely related entities.
Summarise
A model writes a summary for each community ahead of time, so broad questions can be answered from summaries rather than raw text.
How it answers questions
The GraphRAG documentation lists four query modes. Global search answers whole-corpus questions by mapping the question over community summaries and combining the partial answers. Local search starts from specific entities and walks to their neighbours. DRIFT search mixes the two. Basic search is plain top-k vector search for questions ordinary RAG already handles.
The open-source implementation is MIT-licensed on GitHub, with release v3.3.0 on 2026-10-08.
The cost problem and LazyGraphRAG
Building the index means running a language model over the entire corpus, and again when it changes. Microsoft's own answer, LazyGraphRAG (2024-11-25), skips the up-front summarisation and defers model work to query time. Microsoft states its indexing cost is identical to vector RAG and 0.1% of full GraphRAG.
Where GraphRAG fails, and what it is confused with
Lookup questions
For specific facts, hybrid search with a reranker is cheaper and at least as good. GraphRAG earns its cost on sensemaking questions.
Extraction errors
Entities merged wrongly or relationships missed by the model become part of the index, and every answer built on them inherits the mistake.
Fast-changing data
Pre-written community summaries go stale as documents change, so re-indexing cost recurs.
Not just a graph database
Storing data in a graph database does not make a system GraphRAG. The method is the model-built graph plus community summaries, not the storage engine.
How Motiontic's engineers apply it
Motiontic's engineers test GraphRAG when the client's real questions are about patterns across a corpus, such as recurring complaints or themes in research, and compare it with an agent that searches and summarises iteratively. Indexing cost is estimated on a sample before the whole corpus is processed.
Vector RAG, GraphRAG and LazyGraphRAG compared
| Vector RAG | GraphRAG | LazyGraphRAG | |
|---|---|---|---|
| Index built from | Chunk embeddings | Model-extracted graph and summaries | Lightweight graph, no summaries |
| Indexing cost | Low | High, model reads everything | Same as vector RAG (Microsoft) |
| Strong at | Specific facts | Whole-corpus themes | Both, with work at query time |
| Updates | Re-embed changed chunks | Re-extract and re-summarise | Cheaper re-indexing |
Common questions
What is the difference between GraphRAG and RAG?
Standard RAG retrieves passages similar to the question. GraphRAG first builds a knowledge graph and community summaries from the corpus, which lets it answer questions about the dataset as a whole.
Is GraphRAG open source?
Yes. Microsoft publishes GraphRAG on GitHub under the MIT licence; version 3.3.0 was released on 2026-10-08.
Why is GraphRAG expensive?
Indexing runs a language model over every chunk to extract entities and write summaries. LazyGraphRAG cuts indexing cost to 0.1% of full GraphRAG, by Microsoft's figures, by deferring that work.
When should GraphRAG be used instead of vector search?
When questions span the whole corpus, such as main themes or how topics relate, rather than asking for a specific fact found in one passage.
Related
Retrieval-augmented generation (RAG)
Fetch the right passages at query time, then answer from them with citations.
Agentic search
Retrieval as a tool the agent calls, step by step, rather than a pipeline stage.
RAG vs Agentic Search
One retrieval before the answer, compared with an agent that searches as it works.
Chunking (RAG)
Splitting documents into passages before indexing, which quietly sets retrieval quality.
AI product development
New products and major AI features, built from scratch alongside the client's team.
Sources
- Edge et al., From Local to Global: A Graph RAG Approach to Query-Focused Summarization (arXiv, 2024-04-24)
- Microsoft Research, GraphRAG: unlocking LLM discovery on narrative private data (2024-02-13)
- Microsoft Research, LazyGraphRAG (2024-11-25)
- microsoft/graphrag on GitHub (v3.3.0, 2026-10-08)
- GraphRAG documentation (query modes)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.