Well, because the goal is to locate the exact documents in the training set and remove them, not answer a question...

So you stream the training set through the context window of the LLM, and ask it if it contains the requested document (also in the context window).

The advantage is that it can also detect variations of the document.