System Design: Personalized Chatbot Recommending Pins

Design a chatbot that converses with the user and recommends pins inline. Combines an LLM conversational layer with a Pinterest-style retrieval / ranking stack on the back end. New prompt — exact framing not fully transcribed; treat as a hybrid LLM-agent + recommendation design. Requirements Design a conversational product that disambiguates intent, recommends relevant pins inline, personalizes results, and updates recommendations as the dialogue changes. Clarify latency, safety, persistence, and evaluation requirements before choosing the architecture. Notes Use the language model to interpret and reformulate conversational intent, then call a bounded retrieve_pins(query, filters, user_context) tool backed by candidate generation and ranking. Keep ranking authority in the recommender rather than relying on free-form generation. Personalization enters both the conversation context and ranker features. Separate durable user interests from short-lived session intent and define how users can override either. Stream partial text if generation dominates latency, but set independent budgets for retrieval, ranking, and model inference. Add timeouts and a non-personalized fallback. Evaluate recommendation engagement, conversational task success, unsafe or irrelevant suggestions, latency, and repeated-recommendation rate. Offline relevance alone is insufficient. State retention and privacy rules for conversation history before using it as a feature. Preparation Draw the conversational layer, typed retrieval tool, candidate generator, ranker, and response renderer. Write the tool schema and two failure responses: retrieval timeout and no safe result. Define one offline metric, one online recommendation metric, one conversation-quality metric, and two safety guardrails. Walk a multi-turn example where session intent overrides the long-term profile.