Plain RAG retrieves isolated chunks, so it struggles with questions that need connecting facts scattered across many documents — 'which engineers worked on projects that later failed?' GraphRAG first builds a graph of entities and how they relate, then traverses that structure to assemble an answer, often summarizing whole communities of related nodes. It shines on multi-hop and 'whole corpus' questions where the answer is in the relationships, not any single passage. The cost is the upfront extraction pipeline to turn messy text into a clean graph.
ELI5
Think of GraphRAG as a simple recipe for doing the work better. RAG that retrieves over a knowledge graph of entities and relationships instead of (or alongside) flat text chunks.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.