Answers are grounded in this paper and cite their sources.
This work demonstrates improved food guidance in health-conscious consumers, suggesting a more effective advisory model.
Supporting consumers in making autonomous food choices that are sustainable and nutritionally complete is an increasingly complex task, that must take into account several needs to foster eating habits of health-conscious consumers, while reducing food waste and environmental impact. While generative AI and Large Language Models (LLMs) show promising results in this domain due to their natural language processing capabilities, they suffer from critical limitations, including hallucinations, knowledge gaps, and a lack of handling factual information. To mitigate such limitations, Retrieval Augmented Generation (RAG), which retrieves relevant information from external sources to enhance the capabilities of LLMs, has shown effectiveness in many domains. However, existing RAG approaches typically operate on unstructured text that lacks sophisticated symbolic representations of complex domain knowledge. This work proposes an ontology-enhanced conversational food advisory system, that integrates a modular ontology, named FoCOSA (Food Consumer-Oriented Sustainability-Aware), within several key tasks of a RAG-based system, enhancing LLM reasoning with domain knowledge, while the LLM enhances the interpretation of user requests, thus improving retrieval effectiveness and interaction fluidity. Experimental evaluations demonstrate the efficacy of the approach, and the study concludes with guidelines for selecting appropriate settings for food recommendation scenarios considering the complexity of natural language queries and other contextual factors. • Ontology-enhanced RAG improves food advice with health, activity, and sustainability. • FoCOSA ontology enhances LLM reasoning and retrieval in multiple RAG stages. • Empirical evaluation shows improved response accuracy and fewer hallucinations.
No takes yet. Share an insight, caveat, or question.
Bagozi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: