Standard RAG retrieves text chunks. That works well for many factual questions, but it can struggle when the answer depends on relationships: which service depends on which database, which team owns which API, which incidents share the same root cause, or which people worked on two different projects.

GraphRAG adds a graph layer to retrieval. Instead of treating information only as isolated text chunks, the system represents entities and relationships as a structured network. The graph can follow connections, collect related facts, and bring in context that vector similarity may miss.

This chapter covers knowledge graph basics, Microsoft's GraphRAG approach, hybrid vector-plus-graph retrieval, and the trade-offs that determine whether a graph is worth the extra operational cost.