Ask questions about a PDF and actually trust the answers

Here is the direct answer. To ask questions about a PDF and trust what comes back, do three things: use a tool that cites the passage behind every answer, ask narrow questions before broad ones, and run the trap test below before relying on any tool for real work. Most bad experiences with PDF chat are one of two failures, and both are detectable in about two minutes.

The two-minute trust test

Before an exam term depends on a tool, spend two minutes:

The trap question. Ask something the document definitely does not answer (“what does chapter 3 say about the 2019 Nairobi study?” when there is no such study). A trustworthy tool says it cannot find that in the document. An untrustworthy one writes a confident paragraph anyway, silently switching from your PDF to the model’s imagination. Any tool that fails this is decoration.

The citation tap. Ask a real question, then follow the citation to the page. If the passage says what the answer claims, good. If the tool offers no citation to follow, you have no way to run this test, which tells you what you need to know. We covered how the pipeline under the hood works, and citations are the only part of it you can audit yourself.

Questions that work, questions that fail

PDF chat retrieves a handful of relevant passages and answers from those. Questions that map cleanly onto passages work brilliantly: definitions, what-does-the-author-say-about-X, where-is-Y-discussed, summarise-section-Z. Questions that require holding the whole document in mind at once work worst: compare chapter 2 with chapter 7, list every mention of a term, what is the overall argument of this 300-page book.

The fix is decomposition. Ask about chapter 2, ask about chapter 7, then ask for the comparison of the two answers. Ask for the argument of each part before the argument of the whole. You are working with the grain of how retrieval works, and the answers improve immediately.

Three more habits that pay: name locations when you know them (“in the methods section, what was the sample size?”), ask one thing per question, and when an answer surprises you, ask “where exactly does it say that?” before believing it.

Scans, photos and handwriting

A large share of student PDFs are not really text: photographed handouts, scanned chapters, handwritten notes. These need OCR before any of the above works. If your tool returns emptiness or gibberish for a scan, that is an extraction failure, not a model failure, and no prompt will fix it; you need a tool that treats scans as first-class input. Test with a photo of a real page. Tables and figures, meanwhile, survive extraction imperfectly in every tool on the market, so treat any answer read off a complex table as a claim to verify at the source.

Make the answer go somewhere

The question you ask a PDF is usually the start of something: a note, a flashcard, a task. In most chat tools the answer dies in the transcript and next week starts from zero. This is the workflow reason we built Normi the way we did, and yes, it is ours, so verify accordingly: your PDFs and photos live in one library with OCR, every answer arrives with its source pinned, and the same document can become a quiz or a set of reminders without leaving the app. NotebookLM is the strongest free alternative for desk-based interrogation of a source pile, and ChatGPT the strongest for reasoning about what you find; both pass the trust test when set up carefully.

Whichever you use, keep the two-minute test. Normi is on iOS, Android and the web with a 3-day free trial if you want the phone-first version.

If you want the mechanics rather than the technique, asking your own files a question sets out what Normi reads, what it cites and what it refuses to answer.