How to Make a 100 Page PDF Answer Your Questions, and Check That It Is Not Making Them Up

Questo articolo non è ancora disponibile in italiano. Viene mostrata la versione inglese.

What you will have at the end: a long document you can ask questions about in plain language, with an answer that points to a page number, and a thirty second habit for checking whether that page number is real.

An answer looks like this:

Either party may terminate with three months’ notice to the end of a quarter [p. 41].

Thirty seconds later you know whether page 41 really says that.

Not a summary. Summaries of long documents are where people get burned, because a summary is exactly the kind of output you cannot check. This is different. You ask, it answers, you verify, and you move on.

Before you start

• Time: 30 to 45 minutes the first time, about five minutes for every document after that
• Cost: 0 euros with Google’s NotebookLM or the free tier of most assistants. Around 20 euros a month if you want a paid assistant with a larger memory for very long documents
• You need: one PDF and one account with an assistant that accepts file uploads
• Available in the EU: yes. NotebookLM, ChatGPT, Claude and Mistral’s assistant, called Le Chat until May 2026 and now named Vibe, all accept documents here. Mistral processes in the EU by default, which matters if the document is not yours to share
• Worth knowing: NotebookLM’s free tier limits how many questions you can ask per day, which is a real constraint once you get into step 5
• Your data: uploading a document sends it to the provider. Before you upload anything confidential, check the privacy settings and switch off training on your content. On Mistral’s free tier that setting is on by default and has to be turned off. For genuinely sensitive material, a local model on your own machine is the only arrangement where nothing leaves the building

Why this works, in one paragraph

An assistant cannot hold a 100 page document in its head all at once. A model has a context window, roughly the amount of text it can keep in mind at one time, and long documents overflow it. So the tool does something else: it cuts the document into pieces, stores them, and when you ask a question it fetches the pieces that look relevant and answers from those. That is why this method works at all, and it is also the source of every problem below. The tool is not reading your document. It is searching it.

Knowing that changes how you ask.

Step 1: Check that your PDF is actually text

Open the PDF and try to select a sentence with your mouse. Then press Ctrl+F, or Cmd+F on a Mac, and search for a word you can see on the page.

You should now see: the text highlights when you drag across it, and the search finds the word.

If you cannot select anything and the search finds nothing, your PDF is a picture of a document, not a document. That happens with anything scanned or photographed. Skip to the troubleshooting section, because uploading it as it is will produce confident nonsense.

Step 2: Find out how big it really is

Look at the page count. This one number decides your strategy.

For a sense of scale, we ran this on a real 144 page academic survey. It contained about 861,000 characters, which works out to roughly 215,000 tokens, the units models actually count. A useful rule of thumb: four characters make about one token, so you can estimate your own document from its file size or word count.

Many assistants have a context window well below that, especially on free and entry level plans, where 16,000 to 32,000 tokens is common even when the underlying model could handle more.

You should now see: a page count, and an idea of which camp you are in. Under about 50 pages of normal text, most tools will handle the whole thing comfortably. Over 100 pages, assume the tool is searching rather than reading, and ask narrower questions.

Step 3: Upload it

The simplest free route is NotebookLM. Go to notebooklm.google.com and sign in with a Google account. Click “Create new”. A dialog opens asking where the document should come from, with options including Google Drive, a website link and pasted text. Choose the file upload option, pick your PDF, and wait for the processing indicator to finish.

If you would rather use an assistant you already pay for, ChatGPT, Claude and Vibe all work: start a new conversation and attach the file with the paperclip. If you expect to come back to this document repeatedly, put it in a project, a saved workspace that keeps files and instructions between conversations, so you are not uploading it again every time.

Menu names checked in July 2026. These interfaces change often, so look for the paperclip or the words “add source” rather than a specific button position.

You should now see: the file listed as attached or as a source, usually with a page count or file size next to it.

Step 4: Ask for a map, not a summary

Your first message should never be “summarise this”. Ask for structure instead:

List the main sections of this document with their page numbers, and give one sentence per section describing what it covers. Do not summarise the content yet.

You should now see: something that looks like a table of contents with page numbers. This is your navigation, and it doubles as a first test: if the sections it lists do not match the real document, stop and fix the input before going further.

Step 5: Ask real questions, and demand page numbers

Now ask what you actually want to know. Every question gets the same instruction attached:

Answer only from this document. After each statement, give the page number it comes from. If the document does not answer the question, say so instead of guessing.

That last sentence does more work than any other prompt trick. It gives the tool permission to say nothing, which is what you want from a document you intend to rely on.

You should now see: answers with page numbers in brackets, and occasionally a straight “the document does not cover this”, which is a good sign rather than a failure.

Step 6: Verify, in thirty seconds

This is the step everyone skips and the only one that makes the rest trustworthy.

Take one claim from the answer. Note the page number and a distinctive phrase. Open the PDF, jump to that page, and search for the phrase.

Three outcomes:

1. The phrase is on that page, in a paragraph about your topic. Good. Trust the rest a little more.
2. The phrase is on the page, but in a footnote or the bibliography. This is the sneaky one. In our 144 page test document the word “hallucination” appears on 22 pages, but the reference list starts on page 99, so eight of those hits are entries in a bibliography rather than discussion. An answer built on those is technically sourced and practically useless.
3. The phrase is not on that page at all. Stop trusting the answer, and see the troubleshooting section, because the page numbering is probably off.

Do this for two or three claims on your first document. After that you will have a feel for whether a given tool and document behave, and you can spot check instead.

You should now see: the phrase on the page the answer named, in a paragraph that is genuinely about your question. If two out of three checks land like that, the tool is behaving on this document.

Copy this

Paste this as your very first message once the file is attached. It replaces the shorter prompt in step 4 and carries the page number rule from step 5 with it, so you only have to paste once:

You are working only from the document I attached.

Rules:
1. Answer only from this document. Do not use outside knowledge.
2. After every statement, give the page number in brackets.
3. If the document does not answer something, write "not covered in this document" instead of guessing.
4. If a passage is ambiguous, quote it and say what is unclear.
5. Quote exactly when the wording matters, for example deadlines, amounts, obligations and exclusions.

First task: list the main sections with their page numbers, one sentence each.
Then wait for my questions.

When it goes wrong

The PDF is a scan and the answers are invented. This one did not happen in our test run, because the document we used had a proper text layer, but it is the most common failure people report. Symptom: the assistant describes a document that sounds vaguely right but matches nothing. Cause: there is no text layer, only pixels. Fix: run text recognition first. On a Mac, Preview can do it. On Windows, the free LibreOffice Draw or the built in Microsoft Lens can, and Adobe Acrobat does it properly if you have it. Avoid uploading a confidential scan to a random free website, which defeats the point of reading the privacy note at the top. Some assistants, NotebookLM and Gemini among them, will also read a scan as an image. Test the result the way you did in step 1 before you continue.

The page numbers do not match. Symptom: the cited phrase is on a different page, usually by a consistent offset of two or three. Cause: PDF page 5 is often printed page 1, because of covers and front matter. Fix: tell the assistant which one to use. “Use the PDF page position, not the printed page number” fixes it in one line.

The answer misses something you know is in there. Symptom: “not covered in this document” for something you can find yourself. Cause: on very long documents the tool searched and did not retrieve the right piece. Fix: narrow the question. Instead of “what are the termination conditions”, try “look at the section on termination between pages 40 and 55 and list every condition”. Pointing at a region turns a search problem into a reading problem.

Make it yours

Simpler. For a document under 20 pages, skip the notebook entirely. Attach it to a normal chat and ask your question with the page number rule. Two minutes.

Stronger. Load several related documents into one notebook, an old contract and the new one, or three suppliers’ terms, and ask comparison questions: “where do these two documents differ on notice periods, with page numbers from each”.

For a different job. The same six steps work on a manual, a tenancy agreement, a set of meeting minutes, a scientific paper, or a 300 page insurance policy. Only step 2 changes, because the bigger it is, the narrower your questions need to be.

What to try next

If the privacy paragraph at the top made you pause, the next step is doing this without sending anything anywhere. We tested how far that gets you on ordinary hardware in our review of running a local model on a six year old laptop. Short version: for reshaping text it works, and it costs nothing.

Sources

• Anthropic support: upload files to Claude (explains that Claude counts PDF pages as the viewer shows them, which is troubleshooting problem two)
• Google NotebookLM
• Mistral help centre: Le Chat is now Vibe (renamed 28 May 2026)
• A Survey of Large Language Models, arXiv (the 144 page document used to test the numbers in this guide)