How do you implement AI-powered search and discovery in applications?

AI-powered search goes beyond keyword matching to understand user intent and deliver relevant results. Architecture: query understanding (intent classification, entity extraction, spell correction, query expansion), retrieval (hybrid search combining keyword BM25 with vector similarity), ranking (ML-based relevance scoring), and presentation (result formatting, snippet generation, faceted filters). Implementation: index content with both text and vector representations using a search engine that supports hybrid search (Elasticsearch with vector search, Typesense, or PostgreSQL with pgvector + tsvector). Query pipeline: process user query through NLP (extract filters, understand intent), execute hybrid search, apply ML reranker, and return results with highlighted snippets and facets. Semantic search: embed documents and queries using the same embedding model, retrieve by vector similarity for meaning-based matching. This catches synonyms, paraphrases, and conceptual matches that keyword search misses. Personalization: incorporate user behavior (click history, purchase history, browsing patterns) into ranking, implement user embedding that captures preferences, and combine content relevance with personalization signals. Features: autocomplete (trie-based with popularity weighting), "did you mean" (edit distance + frequency-based correction), related searches (query log mining for co-occurring queries), and federated search (search across multiple content types: products, articles, FAQs, people). Evaluation: measure NDCG (quality of ranking), click-through rate (user engagement), zero-result rate (queries returning no results), and mean reciprocal rank (position of first relevant result). Optimization: A/B test ranking algorithm changes, analyze zero-result queries to identify content gaps, monitor search performance by query category, and implement search analytics dashboards.