What it is
GraphRAG uses a language model to extract entities and relationships from your documents, builds a knowledge graph, groups it into communities and writes a summary of each. Queries can then draw on those summaries to answer broad questions, such as the main themes across a whole corpus, that plain vector search handles poorly. Since August 2026 the README has said the project is largely in maintenance mode: Microsoft will fix bugs, update dependencies and patch CVEs, but will not accept new pull requests or add features.
It offers global search (over community summaries) and local search (around specific entities), among other modes. Microsoft Research has also described LazyGraphRAG, a cheaper variant that defers most model work to query time.
It is a Python library and CLI, released by Microsoft Research as demonstration code; the README says it is not an officially supported Microsoft offering.
Our review
GraphRAG is now a finished research artefact rather than a growing library. In August 2026 Microsoft added a notice to the README saying that frontier models had changed sharply since the first release in July 2024 and that the project would get bug fixes, dependency updates and CVE patches but no new features. The releases bear that out: v3.3.0, on 8 October 2026, contained a single SQLite performance fix. The repository has about 36,400 stars.
The idea itself has not dated. Global search answers from community summaries written at index time, so a question like "what are the top themes in this story?" draws on the whole corpus instead of the handful of chunks a vector search retrieves. The quickstart shows it on a Project Gutenberg text, and indexing that one book takes a few minutes with an OpenAI or Azure key.
Cost is the reason most teams stop. The README and the docs both warn that indexing is expensive and advise starting small with cheap models, because every document passes through the model for extraction and summarisation. Microsoft Research's LazyGraphRAG post puts vector RAG indexing at 0.1% of the cost of full GraphRAG. Changed documents largely mean a rebuild, and the README warns that out-of-the-box results may disappoint without its prompt-tuning step.
Upgrades carry their own chores. Minor versions require `graphrag init --force`, which overwrites your configuration and tuned prompts, and major versions need a migration notebook to avoid re-indexing. For a fixed archive indexed once and queried for years, a pipeline that will no longer change is arguably a virtue; for a living document set, LightRAG's cheaper indexing is the comparison most 2026 write-ups reach for.
Where it earns its place
A research analyst at a policy institute with a fixed archive of published reports
Find the themes that recur across the whole collection
Index once with an inexpensive model, tune the prompts, then use global search. The cost is the model tokens for one pass over every report; because the archive does not change, the expensive rebuild never comes round again.
A data engineer at a law firm evaluating retrieval approaches
Prove whether a graph beats hybrid search on cross-document questions before paying for it
Run the quickstart, then index a small sample of matters and compare answers against an evaluation set. Every document goes to the configured OpenAI or Azure endpoint, so the firm's data agreement with that provider applies.
An ML platform team inside an Azure tenant
Offer a graph-based search service without storing API keys on disk
Use azure_managed_identity authentication and keep the parquet outputs in the tenant. With upstream in maintenance mode, the team should expect to own any feature work in a fork.
Before you commit
- Setup
- `python -m pip install graphrag` on Python 3.10 to 3.12, `graphrag init`, add an OpenAI or Azure key, then `graphrag index`; the sample book indexes in a few minutes, per the docs.
- Integrations
- CLI and Python library; the quickstart covers OpenAI and Azure OpenAI, including Azure managed identity. The index is written as parquet files.
- Data and privacy
- Every document is sent to the configured model endpoint for extraction and summarisation; the index stays on local or your own storage. MIT licence; no security advisories published on GitHub.
- Cost at scale
- No paid edition. The bill is model tokens for extracting and summarising the full corpus, repeated when documents change; Microsoft Research puts vector RAG indexing at 0.1% of that cost.
- Limits
- Maintenance mode with no new features; Python 3.13 is not supported; minor upgrades overwrite configuration; changed documents largely force a rebuild.
What's good
- Global search answers whole-corpus questions that vector RAG misses
- Community summaries are useful outputs in their own right
- MIT licence; runs against your own Azure tenant
- Still receives bug fixes and CVE patches
What's not
- In maintenance mode since August 2026; no new features
- Indexing sends the whole corpus through a language model
- Minor upgrades overwrite config and tuned prompts
- Default prompts need tuning, per the README
Best for
Teams with a fixed, high-value document set where cross-document questions justify a costly one-off index.
Alternatives
GraphRAG: common questions
- Is GraphRAG worth it?
- Still the reference design for whole-corpus questions, but Microsoft has stopped developing it and the indexing bill is as large as ever. We rate it 3.5 out of 5.
- How much does GraphRAG cost?
- Free and open source under the MIT licence. Indexing runs on your own OpenAI or Azure OpenAI credits, which the README warns can be expensive.
- What are the downsides of GraphRAG?
- In maintenance mode since August 2026; no new features; indexing sends the whole corpus through a language model; minor upgrades overwrite config and tuned prompts; default prompts need tuning, per the README.
- Who is GraphRAG best for?
- Teams with a fixed, high-value document set where cross-document questions justify a costly one-off index.
- How much does GraphRAG cost for a team or in production?
- No paid edition. The bill is model tokens for extracting and summarising the full corpus, repeated when documents change; Microsoft Research puts vector RAG indexing at 0.1% of that cost.
- Is GraphRAG safe for company data?
- Every document is sent to the configured model endpoint for extraction and summarisation; the index stays on local or your own storage. MIT licence; no security advisories published on GitHub.
- What are the best GraphRAG alternatives?
- The closest alternatives we have reviewed are LlamaIndex, RAGFlow and Haystack.
Answers checked Oct 23, 2026.