Graph Retrieval-Augmented Generation (GraphRAG) can link evidence that is scattered across a corpus graph, yet most systems apply a largely shared exploration procedure to every query. This creates a structural mismatch: direct facts often need compact local neighborhoods, comparison tasks require balanced coverage of multiple targets, and mediated questions may demand deeper paths through weakly related connectors.
Mosaic is a training‑free framework that formulates GraphRAG retrieval as a per‑query control problem. An LLM analyzer translates query‑specific evidence requirements into a bounded policy that decides seed selection, graph traversal, stopping criteria, and final evidence selection, while the corpus graph, indexes, scoring functions, grounding process, and answer generator remain shared.
On the GraphRAG‑Bench, Mosaic achieves a query‑weighted Answer Correctness of 76.97 on the Medical subset and 64.33 on the Novel subset, improving over the previously strongest overall results by 5.13 and 4.43 points respectively. On Medical, Evidence Recall reaches 95.1 and Context Relevancy 86.1. Controlled comparisons on an identical graph and generator show that no fixed narrow, medium, or wide policy is consistently optimal; Mosaic improves by 9.96 points over the strongest canonical fixed policy. Compared to Fixed Wide, it evaluates 81.9% fewer paths and retains 47.2% fewer evidence items.
Transfer experiments on HotpotQA, MuSiQue, and 2WikiMultiHopQA further demonstrate that the policy interface can be applied directly to new benchmarks without benchmark‑specific retriever training.
Review: Mosaic highlights the substantial benefits of query‑aware retrieval policies across diverse tasks, validating the practicality of a control‑oriented perspective.