With smart tourism shifting from “search-price comparison” to “search-and-order,” platforms must respond within milliseconds to dynamic contexts like weather and traffic. Using 2.4 million user logs and 20,000 questionnaires, the authors propose a three-level architecture: (1) a multi-granularity Transformer-Encoder unifies long-term interests and short-term intent via spatiotemporal attention;(2) a gradient-aligned distillation layer compresses high-dimensional sparse context into 512 dimensions, achieving 42ms latency with 97% entropy retention; and (3) a Pareto-aware Contextual Bandit dynamically balances CTR, conversion, and merchant fairness. Experiments show that the new framework is improved by 12.4% and 9.7% on NDCG@10 and MAP@20, respectively. The 14-day online A/B test shows that CTR is improved by 15.3%, the order conversion rate is improved by 9.8%. The recall rate of cold start attractions can still be maintained at 0.43. This study provides a low-cost and portable paradigm and lays the foundation for real-time context modeling of smart tourism.