GraphRAG Knowledge-Graph Retrieval

When an answer requires stitching several documents together, vector search alone falls short: GraphRAG first extracts a knowledge graph, builds a community hierarchy and generates summaries, then retrieves against that structure — strong at whole-corpus questions that need a global view

Tool Interface

Interactive tool will be available soon

Features

✓ Graph first: entities and relations are extracted from raw text into a knowledge graph instead of being chopped into anonymous text chunks
✓ Community hierarchy and summaries: network analysis builds a hierarchy of communities, each summarised, so retrieval can take a higher vantage point
✓ Multiple query modes: global search answers macro questions across a corpus, local search answers detail around an entity, plus modes such as DRIFT
✓ Built for complex reasoning: officially it shows substantial improvement on questions requiring reasoning over complex information versus baseline semantic search
✓ Open source and self-hostable: code and docs are public and run locally with configurable language models, at the cost of expensive indexing — officially, start small

How to Use

Diagnose your question mix: if many ask what a corpus as a whole is about, the graph route is worth trying; if you mostly fetch similar snippets, plain vector search may be cheaper
Index a small slice first: the official warning is that indexing is expensive, so run a handful of documents end to end and check whether the extracted entities and relations make sense
Match the query mode to the question: global search for macro questions, local search for detail around a specific entity — do not use the wrong one
Evaluate and tune: use question generation to build a test set, compare graph-based answers against a vector baseline, then decide whether the indexing cost is justified

FAQ

What is GraphRAG?

A more structured take on retrieval-augmented generation than naive semantic search. Officially, the GraphRAG process extracts a knowledge graph from raw text, builds a community hierarchy, generates summaries for those communities, and then leverages those structures for RAG tasks. It answers questions about a corpus as a whole rather than just fetching similar snippets. See https://microsoft.github.io/graphrag/

How does it differ from plain vector retrieval?

Structure. The common approach chunks documents and does vector similarity, which is good at finding text that resembles the question but often weak on questions needing cross-document synthesis and reasoning. GraphRAG extracts a graph of entities and relations and generates community summaries; officially this structured approach shows substantial improvement on reasoning about complex information. See https://microsoft.github.io/graphrag/ and https://www.microsoft.com/en-us/research/project/graphrag

Which query modes exist?

The docs list global search, local search and DRIFT search, each with worked examples. Global search targets macro questions across the whole corpus, local search zooms in on detail around a specific entity, and DRIFT extends local retrieval with more context and follow-up reasoning. Pick the wrong mode and you tend to get plausible-but-wrong answers. See https://microsoft.github.io/graphrag/

What is the biggest practical trap?

Indexing cost. The repository carries a prominent warning that GraphRAG indexing can be an expensive operation, telling you to read the docs to understand the process and costs and to start small. Extracting entities, relations and summaries with language models across a corpus does not come free, so validate extraction quality on a small slice before scaling up. See https://github.com/microsoft/graphrag

Is the project still actively developed?

One caveat: the official repository states the project is largely in maintenance mode, no longer accepting new PRs or implementing new features, performing only bug fixes and dependency updates, particularly for CVEs. So it remains usable and maintained, but do not expect a stream of new capabilities; a community organisation maintains related GraphRAG resources and forks in parallel. See https://github.com/microsoft/graphrag and https://github.com/graphrag

When should you not use GraphRAG?

When your queries mostly mean find a passage similar to the question, when the corpus is large but queries are narrow, or when budget cannot absorb repeated indexing, plain vector search is usually the better deal. GraphRAG pays off where cross-document synthesis, relational reasoning and a global view are needed; in the wrong setting you spend a lot and gain little. See the official comparison at https://microsoft.github.io/graphrag/