Mosaic addresses a structural mismatch in GraphRAG systems, where uniform exploration procedures are applied across all queries. The framework formulates retrieval as a per-query control problem. An LLM analyzer generates a bounded policy for seed selection, graph traversal, stopping, and evidence selection, based on the specific query’s needs. The corpus graph, indexes, scoring functions, and answer generator remain shared resources. Evaluation on GraphRAG-Bench shows Mosaic achieving a query-weighted Answer Correctness of 76.97 on Medical and 64.33 on Novel, surpassing previous results by 5.13 and 4.43 points respectively. On Medical, it achieves 95.1 Evidence Recall and 86.1 Context Relevancy. Comparisons demonstrate that a fixed, narrow, medium, or wide policy is not consistently optimal, and Mosaic improves over the strongest canonical fixed policy. The policy interface can be applied without benchmark-specific retriever training, as demonstrated in transfer experiments on HotpotQA, MuSiQue, and 2WikiMultiHopQA.
Source: https://arxiv.org/abs/2609.11065