GraphRAG
RAG augmented by a knowledge graph, formalized by Microsoft Research. Instead of chunking your documents and retrieving the closest ones by vector similarity, GraphRAG has an LLM extract entities and their relationships, builds a graph, then answers questions by traversing it. The payoff: it can answer the cross-cutting questions ("what themes recur across the whole corpus?") that flat vector RAG misses, because no single chunk holds the answer.
Strengths
- Answers cross-cutting global questions that flat vector RAG misses by design
- The entity/relationship graph makes reasoning explainable: you see which links the answer traversed
- Open-source reference implementation by Microsoft Research, documented and reproducible
Limitations
- Upfront LLM entity extraction is heavy on tokens and time: high indexing cost
- Rebuilding the graph on every corpus update is costly: a poor fit for fast-changing data
- Overkill for most RAG cases: if the question is local, flat vector RAG is enough
Best for
- Large stable corpora where you ask global synthesis questions (reports, archives, document bases)
- Data/AI teams whose flat vector RAG already fails on measured cross-cutting questions
- Tech leads evaluating when a knowledge graph justifies its overhead versus classic RAG