A Human-Centred Framework for Retrieval-Grounded Cold-Start Recommendation

Cold-start remains a fundamental limitation of recommender systems, particularly when no interaction history is available at first use. In knowledge-intensive domains, this issue is further exacerbated by the inability of existing approaches to exploit rich semantic and contextual information available at first interaction. In this paper, we address this limitation by reframing cold start as a knowledge orchestration problem and proposing a human-centred framework that integrates natural-language preference elicitation, retrieval-grounded candidate generation, and LLM-based reranking with explanation. The framework models recommendations as promptmediated, user-correctable interactions, enabling structured interpretation of first-session input and iterative refinement. The approach is designed to be domain-agnostic and applicable to contexts such as education, healthcare, and cultural heritage. For evaluation, we adopt a widely used benchmark dataset as a proxy testbed, enabling controlled comparison with standard baselines. Results show that combining retrieval grounding with explainable and interactive mechanisms improves first-session recommendation quality and user transparency. This work contributes a conceptual reframing of cold start, a transferable framework, and design insights for human-centred recommendation in semantically rich, behaviourally sparse settings.