AI-Powered Virtual Sommelier: Automated employee training, assessment and wine pairing for the foodservice industry

The high staff turnover and specialized knowledge requirements of hospitality make it difficult to maintain consistent, high-quality wine service. This study investigates whether an AI framework grounded in curated enological knowledge can support staff training, competency assessment, and customer-facing wine advisory within a single workflow. We present the AI-Powered Virtual Sommelier , an integrated system that combines tutoring, assessment, advisory support, and performance reporting. Its main technical contribution is a fixed multi-stage Retrieval-Augmented Generation (RAG) advisory architecture that combines structured metadata filtering, semantic retrieval, bounded LLM-based candidate selection, and expert-grounded response generation within a modular design. The Canary Islands Wine Knowledge Base serves as a regional testbed because it offers curated sensory profiles and pairing guidelines, enabling auditable grounding beyond generic consumer wine datasets while preserving transferability to other wine regions. The proof-of-concept evaluation comprised 108 structured Human-in-the-Loop (HITL) interactions. In this sample, didactic retrieval (10/10), item generation (9/9), and automated grading (9/9) showed no observed failures. Within the advisory suite, semantic recommendation and out-of-scope handling each passed 9/10 cases, whereas ambiguity resolution remained weak (1/10) and the latency protocol did not satisfy its real-time target. Overall, the study shows that an integrated, grounded workflow can operationalize specialized domain knowledge for hospitality, although clarification behavior and response speed still need improvement.