Ontology-Enhanced RAG Architecture for Sensory-Aware Food Recommendation

Food recommender systems must integrate heterogeneous, multi-level knowledge, including nutritional, environmental, and sensory aspects. However, sensory perception, despite its key role in user satisfaction, remains underinvestigated. In this work, we propose a web-based food advisory architecture that includes a Sensory Analysis Ontology, aligned with ISO standards, and embedded within a modular Retrieval-Augmented Generation (RAG) pipeline. Our main contribution is the design and engineering of a web-based system that combines symbolic and sub-symbolic components to enable: (i) ontology-driven semantic indexing of food-related data, (ii) query rewriting using domain-aligned vocabulary, and (iii) a hybrid retrieval layer combining dense vector search with Text-to-SQL over structured sensory databases. This design enables recommendations that match user’s taste preferences based on certified panelists evaluations. Experimental evaluation, performed using a benchmark of real-world, multi-domain queries across nutrition, sustainability, and sensory domains, shows that full integration of ontological knowledge across indexing, retrieval, and generation phases improves response quality, achieving up to +34.6% in actionability, +20.1% in relevance and +26.0% in completeness compared to the baseline.