Retrieval-Augmented Generation (RAG) has been widely adopted to improve the factual accuracy of large language models by incorporating external knowledge. However, traditional RAG systems commonly rely on single-hop or top-k ranking strategies, which are often insufficient for complex queries requiring multi-hop reasoning, particularly in domain-specific scenarios such as university policy consultation. In addition, the high computational cost of large language models limits their practicality for local and low-cost deployment in real-world campus systems.