RAG in One Pipeline

An LLM's parametric memory is stale (frozen at training), uneditable (you can't patch one fact), and uncitable (it can't tell you where it learned something). RAG fixes all three by making context dynamic: at request time, retrieve relevant chunks from your corpus and put them in the prompt — the model reads what it needs instead of recalling what it memorized.

The pipeline shape never varies: query → embed → retrieve candidates → re-rank → assemble context → generate. What varies is which stage is broken on any given day.