GraphRAG

GraphRAG is an extension of retrieval-augmented generation (RAG) in which a knowledge base is indexed not only as vector-embedded text chunks but as an explicit knowledge graph of entities and relations. A language model extracts nodes (concepts, people, terms) and edges (relationships) from documents; retrieval then walks this structure and feeds the model multi-hop paths instead of isolated snippets. The term was popularized by Microsoft Research (Edge et al., 2024). Compared to classic RAG, GraphRAG offers two main benefits: relationships across multiple documents can be traced (for example, common causes mentioned in several quarterly reports), and answers remain auditable because the nodes and edges used are returned as an explicit path. The added structure also costs more: the graph must be built and maintained, and community-based summarization is heavier than simple vector top-k retrieval. In practice, GraphRAG pays off for strongly interconnected sources such as technical documentation, legal files, or master-data models, while small flat corpora are often answered faster and cheaper with plain RAG. Production systems typically combine both: a vector index for fast similarity search plus a knowledge graph for relational queries, coordinated by a router. GraphRAG is therefore less a single tool than an architectural pattern for structured retrieval in AI agents.