AI News

Graph Retrieval-Augmented Generation (GraphRAG) can connect evidence spread across a corpus graph, but systems that use mostly shared exploration procedures may not match the needs of different queries. Direct factual questions may benefit from compact local neighborhoods, comparison questions from balanced coverage of multiple targets, and mediated questions from deeper paths through weakly related connectors. MOSAIC addresses this mismatch as a per-query control problem and does not require training a retriever. An LLM analyzer converts each query's evidence requirements into a bounded policy covering seed selection, graph traversal, stopping, and evidence selection, while the corpus graph, indexes, scoring functions, grounding procedure, and answer generator remain shared. On GraphRAG-Bench, it reports query-weighted Answer Correctness scores of 76.97 on Medical and 64.33 on Novel, exceeding the strongest previously reported overall results by 5.13 and 4.43 points. On Medical, it also reaches 95.1 Evidence Recall and 86.1 Context Relevancy. Controlled tests using the same graph and generator show that no fixed narrow, medium, or wide policy is consistently best; MOSAIC improves by 9.96 points over the strongest canonical fixed policy. Compared with Fixed Wide, it evaluates 81.9% fewer paths and retains 47.2% fewer evidence items. Transfer experiments on HotpotQA, MuSiQue, and 2WikiMultiHopQA indicate that the policy interface can be reused without benchmark-specific retriever training.