Standard RAG retrieves documents by matching question vectors to chunk vectors, working well for local queries like "Which service owns the payments retry logic." However, global queries such as "Which failure causes recur most often across all postmortems" require reasoning across entire datasets.

Graph RAG addresses this limitation by building knowledge graphs that capture entities (services, teams, incidents) and relationships between them. This allows answers to emerge from corpus-wide patterns rather than single documents.

Microsoft tested Graph RAG against standard vector retrieval with 8,000 and 64,000 token contexts. On global questions, larger windows still failed to match Graph RAG's comprehensiveness, diversity, and source quality.

LinkedIn's customer service team reported a 77.6% improvement in mean reciprocal rank and 28.6% reduction in median resolution time after implementing knowledge graph-based retrieval. Graph RAG maintains two graphs: a lexical graph linking documents to chunks, and an entity graph connecting described concepts.