Basic vector search is a good baseline, but it is rarely the final retrieval system.

A user asks, "What is the refund policy?" The relevant document says, "Customers may return physical products within 30 days of delivery." The meaning is close, but the wording is different. Another user asks about "SSO setup," while the docs say "SAML identity provider configuration." A third asks for an error code, where semantic similarity is less useful than exact keyword match.

These are retrieval problems, not generation problems. If the right evidence never reaches the model, prompt engineering will only hide the failure.

This chapter covers the retrieval techniques that usually matter after the first RAG prototype: query expansion, HyDE, hybrid search, reranking, parent-child retrieval, contextual retrieval, multi-step retrieval, and structured-data routing.