The document's text is extracted and sent to a language model. Anything longer than the model's context is split, summarised piecewise and merged, so a two-hundred-page report is handled the same way as a two-page one.