Although Retrieval-Augmented Generation (RAG) empowers Large Language Models (LLMs) with external knowledge, current frameworks struggle with complex multi-hop reasoning as they fundamentally rely on shallow retrieval signals. Existing methods typically employ coarse semantic matching or entity-centric graph expansion, which often retrieve textually similar but factually irrelevant evidence, injecting noise into the reasoning process. To overcome this limitation, we introduce FACTA, a structured fact alignment framework for multi-hop retrieval. FACTA bridges the gap between a query’s step-wise information demand and a document’s information supply by translating both into structured, explicit factual triples. Through a dual mechanism of triple-level fact matching and multi-signal document reranking, FACTA better identifies evidence that satisfies the factual prerequisites of each reasoning hop. Extensive evaluations on three multi-hop QA benchmarks show that FACTA substantially outperforms state-of-the-art baselines, demonstrating that structured fact-centric retrieval effectively reduces reasoning noise and enhances the accuracy of multi-hop generation.