GraphRAG

GraphRAG is Microsoft's open-source framework for retrieval-augmented generation over knowledge graphs. Instead of chunking documents and hoping vector search finds the right passages, it uses an LLM to build a rich knowledge graph from your documents — entities, relationships, claims — then answers questions by traversing that graph, including global questions about whole datasets ('what are the main themes across these reports?') that vanilla RAG fumbles.

The pipeline runs in stages (indexing: extract graph, build communities, summarize; querying: local search over entities or global search over community summaries), and it is highly tunable for quality-vs-cost tradeoffs. With ~36,300 GitHub stars it is one of the most influential RAG architectures, and the ideas have spread across the ecosystem.

The honest caveat: indexing is expensive — building the graph costs significant LLM tokens, and it is a batch pipeline, not a live index. For simple Q&A over a few documents, standard RAG is cheaper and faster; GraphRAG earns its keep on large corpora where connections between documents matter.

1. pip install graphrag (Python 3.10+).
2. graphrag init --root ./myproject, drop documents into the input/ folder.
3. Set your LLM API key and configure settings.yaml (model, chunking, community settings).
4. Run graphrag index --root ./myproject, then query with graphrag query --root ./myproject --method global "your question".