ChatGPT cites its sources. You see the neat little [1] , [2] markers, and the implicit message is: the model went out, looked at the web, brought back evidence, and is showing you receipts. That story is half right. The other half is what every team building a RAG system gets wrong. There is no single retrieval system inside ChatGPT. There are at least two — a parametric one frozen in the weights and a live one that fires only sometimes plus a tool layer deciding which to invoke, plus a generation step that has to reconcile them when they disagree. Almost none of it is published in detail. Some is confirmed by OpenAI and Microsoft. Some is inferred from leaked system-prompt fragments and citation studies. A lot is just observable behavior if you poke it with enough queries. I spent a week tracing the pipeline. What follows is an engineer's reading of how it actually works the two channels, the eight-step pipeline, the tool layer, and the one finding that should change how you build your own retrieval system.…