# Design User and Item Embeddings from Long Histories Design an embedding system that represents users and visual-content items for retrieval or recommendation. User histories may contain up to 160000 events. ### Constraints & Assumptions - Practice assumption: events include impressions, saves, clicks, hides, and timestamps. - Items have image and metadata features; new items may have little interaction history. - The online retrieval path has a strict latency budget and serves millions of users. - Training examples must respect event time to avoid future leakage. - The long-history limit is a prompt constraint, not a claim about any production implementation. ### Clarifying Questions to Ask - Is the embedding used for candidate retrieval, ranking features, or both? - What action and time horizon define relevance? - How fresh must user and item representations be? - Which privacy, retention, and deletion rules constrain history? ### Part 1: Objective and Training Data Define examples, positives, negatives, loss, leakage controls, and offline evaluation. #### Hints - The way an item was exposed affects what a missing interaction means. #### What This Part Should Cover - Objective aligned with serving use - Defensible negative sampling - Time-correct splits and metrics ### Part 2: Model and Long-History Representation Choose user and item encoders and explain how you preserve useful signals from as many as 160000 events without processing all events naively on every request. #### Hints - Different time scales may deserve different summaries. - A representation can be updated incrementally. #### What This Part Should Cover - Multimodal item representation - Efficient long-sequence strategy - Freshness and cold-start behavior ### Part 3: Serving, Evaluation, and Safety Design embedding generation, storage, nearest-neighbor retrieval, refresh, rollout, monitoring, and deletion. #### Hints - Training-serving skew can occur in both features and index versions. #### What This Part Should Cover - Versioned online architecture - Latency and freshness trade-offs - Online metrics, guardrails, and privacy operations ### What a Strong Answer Covers - A clear target and unbiased evaluation strategy - A scalable approach to very long histories - Cold-start and multimodal item handling - Versioned retrieval serving with monitoring and deletion guarantees ### Follow-up Questions - How would you preserve multiple user interests in one retrieval request? - How would you detect popularity collapse in the embedding space? - What changes if event history must be deleted within minutes? - When would a sequence model outperform engineered summaries enough to justify its cost?