Large language models generate fluent text. They fail to meet grounding, traceability, freshness, and access control requirements. Retrieval-Augmented Generation addresses this by forcing models to answer using external evidence.

Early RAG used one simple pipeline. Production systems now use multiple architecture patterns. Each pattern targets a...