GraphRAG (Graph Retrieval-Augmented Generation) is an extension of standard RAG that organises retrieved knowledge as a graph of entities and relationships rather than a flat list of document chunks. A knowledge graph is first built from the source corpus — entities (people, concepts, organisations) become nodes, and stated relationships become edges. At query time, the system traverses this graph to collect multi-hop evidence: facts that are not in any single document but emerge from chaining connected nodes. Microsoft Research published the canonical GraphRAG paper in 2024, demonstrating substantial gains on questions that require synthesising information across many sources.

Think of it like the difference between searching a library's card catalogue (standard RAG: find the closest-matching card) versus asking a reference librarian who has read everything and knows how every book, author, and idea connects to every other (GraphRAG: traverse the relationships to build a multi-step answer).

A simplified GraphRAG pipeline looks like this:

1. Ingest documents → extract entities + relations → store as graph (e.g. Neo4j or in-memory NetworkX)
2. At query time: embed query → find seed entities → BFS/DFS over edges → collect N-hop context
3. Summarise subgraph context → pass to LLM as grounded prompt

GraphRAG builds on RAG (the base retrieval-augmentation pattern), Embedding (entities and queries are embedded for similarity search), Vector Database (often used alongside the graph for initial seed retrieval), and Hallucination reduction (multi-hop grounding gives the model more connected, verifiable evidence). It is also related to Knowledge Distillation in that structured knowledge extraction is a prerequisite.

You will encounter GraphRAG in enterprise search, legal and scientific research tooling, and any application where questions require synthesising relationships across a large document corpus — rather than finding a single relevant passage.