Design an ML-powered search system

## Scenario Design an end-to-end search system for a consumer product (e.g., an e-commerce marketplace or content platform) where users type queries and expect relevant, personalized results. ## Requirements ### Functional - Given a query string, return a ranked list of results (items/documents/videos/etc.). - Support: - Keyword matching (lexical search). - Semantic matching (synonyms/paraphrases via embeddings). - Filters/sorts (category, price range, recency, etc.). - Autocomplete/suggestions (optional but preferred). - Handle frequent content updates (new/edited documents should become searchable quickly). ### Non-functional (assume reasonable targets) - Latency: p95 < 200 ms for the online request path. - Scale: 10k QPS peak; corpus size ~100M documents. - Reliability: graceful degradation if ML components fail. - Observability: logging for debugging, offline training, and online evaluation. ## What to cover - High-level architecture (offline pipelines + online serving). - Retrieval (candidate generation) and ranking. - Indexing strategy and freshness. - Model training data, labels, evaluation metrics. - Personalization and cold-start. - A/B testing and launch plan.