Hybrid Search and Reranking

1. Vector Search Alone Ships the Demo, Not the Product

Pure embedding retrieval has a blind spot, and every production RAG team meets it in week one: exact tokens, meaning the literal words, IDs, and codes sitting inside a query. Ask "what does error E4022 mean?" Ask "summarize ticket JIRA-8871." Ask "what's the SLA in the Acme contract?" Semantic similarity just shrugs at questions like these. Here is why. Semantic search means matching by meaning: it turns text into vectors (short lists of numbers) and finds chunks whose vectors land close to the query's vector