Hybrid Search

Hybrid Search is a retrieval strategy that combines traditional keyword scoring (e.g., BM25) with dense vector similarity to surface results that match both exact terms and underlying meaning. In practice, a query first runs through a lexical engine to capture precise phrases and filters, then through an embedding-based semantic index. Scores are combined—via weighted sum, reciprocal rank fusion, or learning-to-rank—to produce a single, relevance-optimized list. This approach boosts recall on synonymy (“doctor” ↔ “physician”) while preserving precision for numeric filters and rare strings like product SKUs. Hybrid Search underpins modern e-commerce, legal discovery, and Retrieval-Augmented Generation (RAG) pipelines, where higher-quality context slashes hallucinations and token costs. Key tunables include weight ratios, rerank depth, and metadata constraints; metrics such as NDCG, recall@k, and latency gauge trade-offs. By blending the strengths of lexical and semantic methods, Hybrid Search delivers search experiences that “think” and “read” simultaneously.