Retrieval over a knowledge graph as well as over text. It shines on questions that require joining facts scattered across many documents, or summarising a whole corpus, which plain chunk retrieval handles badly because no single chunk contains the answer.
Technical view
Extracts entities and relationships into a graph, optionally clusters it into communities with generated summaries at each level, and retrieves subgraphs or community summaries rather than isolated chunks. Strong on multi-hop and global questions; weak on cost, since graph construction requires a model pass over the whole corpus, and on maintenance as documents change. Hybrid designs retrieve both graph context and raw text.
Related
Part of AI Terminology Index, section: Retrieval and Knowledge.