Retrieval‑augmented generation (RAG) hinges on fetching evidence needed for reasoning, yet multi‑hop question answering poses a special challenge because supporting passages are linked through intermediate entities and relations that must be uncovered step by step. Existing retrieval pipelines usually rely on a single intent or a one‑shot query expansion, which limits their ability to adapt to newly retrieved evidence and may introduce noisy or redundant signals.\ \ We introduce a training‑free multi‑hop retrieval framework that combines evidence‑conditioned exploration, passage‑specific contrastive refinement, and coverage‑aware final ranking. During offline indexing, the system builds contrastive facets for each passage, describing how the passage differs from its semantically similar neighbors; these facets provide fine‑grained cues to separate closely related candidates. At inference time, the framework iteratively: (1) retrieves current evidence; (2) generates probes targeting unresolved information needs; (3) refines candidate relevance using the contrastive facets; and (4) selects a complementary set of passages that together cover diverse evidence‑seeking intents.\ \ Experiments on MuSiQue, HotpotQA, and 2WikiMultihopQA demonstrate consistent gains in both retrieval quality and downstream QA performance over strong baselines.\ \ Review