Main idea

Loading PDFs into NotebookLM or into a ChatGPT conversation and asking your questions against them: the tool finds the passages close to the question, assembles them, answers — and then everything is thrown away. The next question starts again from the same documents.

What is missing is not speed. It is the intermediate product. A question that requires cross-referencing five documents forces that cross-referencing to be redone every time, and a cross-reference made in the rush of an answer is not worth the one you would have laid down, reread and corrected. The ideas stay weak in the precise sense that they were extracted one by one from each source, without ever being set against one another.

A synthesis, a link between two concepts, a contradiction spotted between two authors are objects that exist before the question or never exist at all. No assistant manufactures them at the moment you query it.

Why it matters

The diagnosis decides the remedy. Attributed to the model, the weakness of the answers leads you to change models or rewrite your prompts — work with no effect on the cause. Attributed to the absence of any artifact kept between two questions, it leads you to build that artifact, which is a project of another nature.

It also gives a criterion for judging a document-search tool: not what it retrieves, but what it leaves behind once the answer has been given.

Nuances and limits

For a question that fits in a single document, reassembling on the fly is amply sufficient, and consolidating would be waste. The gap only appears on what requires cross-referencing.

And the kept artifact carries the opposite risk: it freezes one reading. The next question then inherits the consolidation you made, rather than the documents themselves.

Open questions

• How do you estimate, before building, what share of the questions to come will really require cross-referencing several sources?