關鍵字：大型語言模型、改良式檢索增強生成、智慧旅遊推薦、倒排問題匹配、意圖對齊、候選類型平衡;With the rapid development of Large Language Models (LLMs)， natural language systems ave become increasingly capable of providing travel information and recommendations based on user requirements. However， relying solely on LLMs may result in outdated information， insufficient supporting evidence， and hallucinations. Retrieval-Augmented Generation (RAG) addresses these issues by retrieving relevant information from external data sources to support response generation.
This study focuses on tourism recommendation in Penghu and proposes an enhanced RAG architecture， QuIM-RAG+. The proposed architecture integrates query enhancement， semantic chunking， candidate fusion， reranking， and candidate type balancing to improve candidate retrieval and ranking for multi-constraint travel queries. Both original and balanced datasets containing tourist attractions， restaurants， and accommodations are used to examine the influence of data type distribution.
QuIM-RAG+ is compared with Baseline RAG， Hybrid RAG， QuIM-RAG， TouaRAG- HyDE， and Chat2Route-inspired methods. The evaluation considers candidate retrieval， ranking quality， recommendation hit performance， grounding， query intent alignment， and candidate type diversity. The results show that QuIM-RAG+ achieves better retrieval， ranking， and recommendation performance under both dataset settings. The candidate type balancing mechanism also reduces excessive concentration of Top-K candidates in a single travel category， indicating that dataset balancing alone cannot completely resolve candidate type concentration.
Overall， QuIM-RAG+ improves candidate retrieval， requirement matching， and candidate type diversity in regional tourism recommendation. The results also show that recommendation quality depends not only on the generation model， but also on the quality of candidate retrieval， ranking， and candidate composition.
