Authors
Ratomir KarlovićMia RovisAlma SmajićL.E. SeverIvan Lorencin
Topics
Diverse Aspects of Tourism ResearchDigital Marketing and Social MediaRecommender Systems and TechniquesArticleRatomir Karlovi´c , Mia Rovis , Alma Smaji´c , Luka Sever and Ivan Lorencin *Faculty of Informatics, Juraj Dobrila University of Pula, Alda Negrija 6, 52100 Pula, Croatia;ratomir.karlovic@unipu.hr (R.K.); mia.rovis@unipu.hr (M.R.); alma.smajic@unipu.hr (A.S.);luka.sever@unipu.hr (L.S.)Correspondence: ivan.lorencin@unipu.hrAbstractThis study evaluates the performance of seven large language models (LLMs) in generatingcontext-aware recommendations. The system is built on a collection of PDF documents(brochures) describing local events and activities, which are embedded into an FAISS vectorstore to support semantic retrieval. Synthetic user profiles are defined to simulate diversepreferences, while static weather conditions are incorporated to enhance the contextual relevance of recommendations. To further improve output quality, a reranking step, utilizingCohere’s API, is used to refine the top retrieved results before passing them to the LLMsfor final response generation. This allows better semantic organization of relevant contentin line with user context. The main aim of this research is to identify which models bestintegrate multimodal inputs, such as user intent, profile attributes, environmental contextand how these insights can inform the development of adaptive, personalized recommendation systems. The main contribution of this study is a structured comparative analysisof 7 LLMs, applied to a tourism-specific RAG framework, providing practical insightsinto how effectively different models integrate contextual factors to produce personalizedrecommendations. The evaluation revealed notable differences in model performance, withQwen and Phi emerging as the strongest performers, whereas LLaMA frequently producedirrelevant recommendations. Moreover, many models favored gastronomy-related venuesover other types of attractions. These findings indicate that although the RAG frameworkprovides a solid foundation, the selection of underlying models plays an important role inachieving high quality recommendations.Keywords: context-aware tourism recommendations; information extraction; large language models (LLMs); machine learning; personalized recommendation systems; recommender systems; retrieval-augmented generation (RAG)