The intelligent operation and maintenance of station-city integration cyberspace represents a cross-domain application of intelligent transportation and smart cities. It involves complex sensor network deployment and semantic fusion of multi-source heterogeneous data. To address this challenge, this paper proposes OmsRAG, a dual-modal knowledge base framework integrating temporal graph neural networks with retrieval-augmented generation. First, a knowledge graph is constructed through ontology modeling of station-city integration space. Static knowledge, including sensor deployment schemes and spatial operational performance indicators, is stored in a graph database. Meanwhile, spatiotemporal features are extracted from monitoring data by graph attention networks and gated recurrent units. These features are used to generate a dynamic vector knowledge base for operational data. Together, these two components form a dual-modal knowledge foundation. On this basis, a query routing-driven dual-channel retrieval-augmented generation framework is designed. Semantic intentions are recognized through prompt templates, enabling intelligent routing between graph query channels and vector query channels. Dual-channel results are then fused to generate comprehensive answers. Experimental results demonstrate that OmsRAG significantly outperforms Naive RAG and Graph RAG baselines in both context recall and precision.