Design a music recommendation system, then explain where an LLM-based conversational recommendation agent adds value without replacing the core ranking system. Cover candidate generation, ranking, feedback, cold start, evaluation, and safety. ### Constraints & Assumptions - The catalog is large and user preferences change by context and time. - Recommendations must respect regional availability and explicit-content settings. - Conversation latency and model cost must be bounded. - The answer should compare a conventional recommender with the agent extension. ### Clarifying Questions to Ask - Is the surface a single next track, a playlist, search, or a conversational session? - Which feedback signals are available and how quickly should they affect results? - What diversity, novelty, and repetition constraints matter? ### What a Strong Answer Covers - Multi-stage retrieval and ranking using collaborative, content, and contextual signals. - Training labels, negative feedback, exploration, cold start, and feedback-loop risks. - A policy layer for availability, safety, diversity, and session constraints. - An LLM that turns conversation into structured preferences and explanations while calling bounded recommendation tools. - Offline metrics, online experiments, guardrails, latency and cost budgets, and rollback. ### Follow-up Questions - How would you prevent the agent from recommending a nonexistent track? - How would you handle a user asking for something similar to a very new artist? - Which online metric would detect that engagement rose because recommendations became repetitive?