GraphRAG is a retrieval-augmented generation approach that builds a knowledge graph of entities and their relationships from your documents, then…
See the question types plain chunk retrieval can't answer well.
Standard RAG retrieves the top chunks most similar to a question and hands them to the model. That works for questions whose answer sits in one passage, but it struggles with two kinds. Multi-hop questions need facts spread across documents — 'which project did the person who founded company X later advise?' — where no single chunk holds the answer.
Global questions need the whole corpus — 'what are the main themes across these 500 reports?' — but similarity search only ever returns a handful of local chunks, so it can't summarize the big picture. GraphRAG targets exactly these gaps by capturing how facts connect.
Turn documents into a graph of entities and relationships.
GraphRAG's indexing does more than embed chunks. An LLM reads the documents and extracts entities (people, organizations, concepts) and the relationships between them, assembling a knowledge graph where nodes are entities and edges are typed connections. The source text stays linked so answers remain grounded.
Many implementations then detect communities — clusters of densely connected entities — and generate a summary for each. These community summaries become a hierarchical map of the corpus that can be queried for high-level, global questions.
Query the graph two ways depending on the question.
GraphRAG answers with two strategies. Local search starts from the entities named in the question, gathers their neighbors and connecting facts by traversing the graph, and answers from that focused neighborhood — ideal for multi-hop questions about specific things.
Global search uses the community summaries: it consults the hierarchical summaries across the corpus and combines them to answer broad, thematic questions no single chunk could. The mode follows the question — pinpointed and connected, or sweeping and thematic.
GraphRAG doesn't have to replace vector search. Many systems combine graph traversal with vector retrieval — using the graph for connections and structure, and embeddings for fuzzy semantic matching — so the two cover each other's gaps.
Weigh GraphRAG's cost against its reach and decide when to use it.
GraphRAG's strengths — multi-hop reasoning, global summarization, and traceable connections — come at a real cost. Building the graph runs an LLM over the whole corpus to extract entities and relationships and to summarize communities, so indexing is far more expensive and slower than plain chunk-and-embed, and it must be refreshed as documents change.
Use GraphRAG when your questions genuinely span many documents or need corpus-wide synthesis, and the domain has rich entity relationships (research, intelligence, complex knowledge bases). For simple fact lookup in one passage, standard vector RAG is cheaper and enough.
Watch for: reaching for GraphRAG when simple retrieval would do (paying big indexing cost for no benefit); trusting an entity graph the LLM extracted without spot-checking its accuracy; and letting the graph go stale as sources update. Extraction quality is the ceiling — a wrong or missing relationship silently breaks multi-hop answers.
GraphRAG builds a knowledge graph of entities and relationships from your documents and retrieves over it, filling standard vector RAG's gaps on multi-hop questions and global, corpus-wide synthesis. Indexing uses an LLM to extract entities and summarize communities; querying uses local search (traverse entity neighborhoods) or global search (combine community summaries), often hybridized with vector retrieval. The reach costs expensive indexing and depends on extraction quality, so reserve it for genuinely connected or thematic questions.
You must answer 'what are the recurring risk themes across 300 incident reports, and which teams are connected to them?' Explain why standard vector RAG struggles, how GraphRAG's global and local search would each contribute, and one check you would run on the extracted graph's accuracy.
What is GraphRAG?
GraphRAG captures how facts connect as a graph, enabling retrieval that traverses relationships rather than returning isolated chunks.
Which questions does GraphRAG handle better than standard vector RAG?
Vector search returns a few local chunks; GraphRAG traverses connections for multi-hop answers and uses community summaries for global themes.
What is the difference between local and global search in GraphRAG?
Local search is entity-focused traversal for specific multi-hop questions; global search uses hierarchical summaries for corpus-wide questions.
What is the main cost of GraphRAG?
Graph construction runs an LLM over the corpus, making indexing costly; it is worth it for multi-hop or global questions, not simple single-passage lookups.