Abstract
AI-driven travel assistants are increasingly adopted in the tourism sector, yet they often suffer from hallucinations, outdated knowledge, and limited personalization. To address these challenges, we introduce TouaRAG, a multi-layer RetrievalAugmented Generation (RAG) framework designed for personalized and context-aware travel assistance. TouaRAG integrates persona-aware query reformulation with multiple retrieval strategies, enabling dynamic adaptation to user preferences and domain-specific knowledge. We systematically evaluate five retrieval configurations, such as BM25, Hybrid Vector Retrieval, AutoMerge, HyDE, and GraphRAG, across both locally hosted and cloud-based large language models using a tourism-specific benchmark. Experimental results show that HyDE consistently achieves the best balance between retrieval coverage and generation quality, while simpler methods offer competitive performance under resource constraints. Our findings highlight the importance of domain adaptation and persona-conditioned retrieval in improving the reliability and relevance of AI-driven travel assistants. TouaRAG is released as an open-source framework to support reproducible research and real-world deployment.
Showing the abstract — retrieve the full paper via the Exa API.