How to work with a long document using AI

Start with the document’s structure

Imagine a fictional, 40-page supplier offer. You need to understand what is included, what costs extra and what to ask the supplier. Skim the contents and appendices, note the file’s version, then ask the AI for a map: main sections, page ranges and the questions each section answers. Use it to locate relevant passages while keeping related appendices in view.

Make sure the input is readable

Try copying a few lines from the file, including a table. For scans, OCR—software that turns pictured writing into searchable text—can introduce errors; Adobe explains how to review uncertain words against the original image. Check numbers and units before building notes from them.

Check whether your tool reads page images or only extracted text. Google’s Gemini API documentation distinguishes PDF vision from text-only treatment of other document types. That behavior is specific to the documented service. Inspect the original chart or table whenever its layout carries meaning.

Ask for a bounded briefing

Name the question and output, for example:

Using only this offer, prepare a one-page briefing on included items, extra charges and delivery conditions. First locate the relevant passages. For every point, give the page and section; mark missing information as unstated. Finish with three questions for the supplier.

Anthropic recommends finding relevant quotations before answering long-document questions. Here, that means gathering the source passages before compressing them into notes.

If it is too long, work by section

If your tool cannot accept the whole document, split it at section boundaries, keeping headings, page labels and related appendices. Use the same question for each part, then combine the notes and check cross-references.

A large input limit does not establish exhaustive reading. In the 2024 Lost in the Middle study, tested models sometimes retrieved information less successfully when it appeared in the middle. Those experiments concern specific models and tasks, not a performance score for today’s tools.

Keep the condition beside the number

In our fictional offer, page 6 lists €25 per unit; page 31 says delivery costs extra. The note should retain both: €25 per unit, excluding delivery (pages 6 and 31). Open those pages and compare the wording, currency and units. If a condition is unclear, keep it as a question for the supplier.