A model knows what it was trained on and nothing else: not your documents, not yesterday's prices, not your project's rules. Retrieval is the step that fetches the right pieces of your own text and hands them to the model as context. It is the difference between a model that guesses and one that reads.
About 1 min of reading
State the contract in one line.
The question
Before you run it, say out loud what it will print. Then open the answer.
The passage that answers it
Before you run it, say out loud what it will print. Then open the answer.
Write the question the system must answer, then find the passage that answers it. If you cannot find a passage that answers it, retrieval cannot help and the document is missing.
Print the passages that were chosen before the model sees them. Most bad answers are a bad retrieval, and reading the passages is the fastest way to see that.
The model answers from the text you put in the prompt. If the answer needs a fact from your files, that fact has to be in the prompt. Retrieval is the machinery that decides which facts go in.
A model with no retrieval writes a plausible answer about your system that is wrong in the details. That fluency is why the failure is invisible, and why the citation of the source you retrieved is worth requiring.
You change what the model reads without changing the model. That is why retrieval updates in a day, works on private documents, and can be checked by reading the passages it chose.