Your Manual Has the Answer. Here's Which AI Can Actually Find It

An amber symbol appears on the dashboard that you have never seen before. You have the manual — a 180-page PDF sitting in a folder somewhere — and searching it for “orange light” returns nothing, because the manual calls it a malfunction indicator and puts it in a table on page 94. This is the ordinary case for using AI to read a product manual, and it is a much better job for a language model than most of the things people try.

It is also the case where the free tools quietly disagree with each other. Not on quality — on whether they look at the pictures at all.

The question that decides which tool works

Every question you will ever ask a manual falls into one of two piles.

The first pile is text. What does this error code mean, how long is the warranty, what oil does it take, which button starts the self-clean cycle. The answer exists as a sentence somewhere in the document and the only hard part is finding it. Every tool here does this well.

The second pile is pictures. Which of these four sockets is the one in the photo, what is part 7 in the exploded diagram, which way does the filter face, what does that symbol on the control panel mean. The answer is in a figure, and a tool that reads only the text of the page will not see it.

That split matters because the tools do not warn you which pile you are in. Ask about a diagram in a document whose images were never processed and you will get an answer anyway — assembled from captions, surrounding paragraphs and a confident guess. It will read exactly like the answers that are correct.

So before anything else: look at the page your answer is probably on. If the information lives in a figure, your choice of tool is already made for you.

Claude, and the 100-page line

Claude’s own documentation draws the sharpest limit of the three, and almost nobody mentions it. According to Claude’s file upload documentation, it “analyzes both text and visual elements (like images, charts, and graphics) in PDFs of 100 pages or fewer. For PDFs from 101 to 1000 pages, Claude processes text only and doesn’t analyze visual elements.”

Read that against your own manual, and the first thing to do is check its page count — your PDF viewer shows it. A short appliance booklet will often sit under the line. A car manual, which is where the diagrams you actually need tend to live, often will not. Above 100 pages, every exploded diagram, every dashboard symbol chart and every wiring figure in the document is invisible to the model, and you will not be told. Over 1000 pages the upload is refused outright with an “Uploaded file is too large” error, which is at least honest.

The rest of Claude’s chat limits are generous: 500MB per file and up to 20 files in a single conversation. Projects, which keep a document available across conversations rather than in one chat, are tighter — 30MB per file, and the documentation describes project files as “text extraction only (except for multimodal PDFs)”. Projects are available on free accounts, capped at five, but the retrieval that expands a project’s capacity is a paid-plan feature.

For a short manual with diagrams that matter, Claude is the best of the three. For a long one, it is the worst, and the failure is silent.

Gemini Notebook, formerly NotebookLM

Google renamed NotebookLM to Gemini Notebook, which is worth knowing because half the guides you will find still use the old name and point at a help centre that has since moved to a different address.

On the free tier it takes up to 50 sources in a notebook, with each source capped at 500,000 words or 200MB for an uploaded file. There is no page limit, which is the headline difference: a 600-page service manual goes in whole. It will also hold the manual, the warranty document, the recall notice and the forum thread you found, all as separate sources in one notebook, and answer across them.

The more useful detail is buried in the same page: images are a supported source type, including JPEG, PNG, HEIC and WebP. That gives you a way around the diagram problem that works everywhere. Photograph the page, or crop the figure out of the PDF, and add it as its own source. The model is then looking at the picture directly rather than at a page of text that happens to contain one.

The failure that happens before you start

There is one way a manual simply will not go in, and it is common enough to check first. Google’s FAQ lists, among the reasons an import fails, that “your original PDF file is copy-protected”.

A copy-protected file opens perfectly well in a PDF viewer, so nothing looks wrong until the upload bounces, which makes this a confusing failure to hit halfway through a job. If it happens, check whether the manufacturer publishes a separate owner’s manual alongside the service document — where both exist, the consumer-facing one is the one to try next.

ChatGPT and the three-uploads-a-day ceiling

ChatGPT handles a manual perfectly well. The constraint is how many times you get to try.

OpenAI’s file upload FAQ puts free accounts at three file uploads per day, with documents capped at 512MB and two million tokens each. One manual is one upload, so for a single question that is plenty. It stops being plenty the moment you are working through a repair, realise the manual you grabbed is for the wrong model year, download the right one, and find you have one upload left before tomorrow.

The way to spend the allowance well is to upload once and keep the manual in a project rather than re-attaching it to each new chat, so the day’s uploads go on documents you do not already have.

How to use AI to read a product manual, step by step

Get the manual from the manufacturer. Not from a manual-aggregator site. An aggregator copy can be the wrong revision for your model year, and a 2019 manual answering a question about a 2023 model is the kind of error you will not catch, because it will be internally consistent and completely wrong.

Check it opens and is not copy-protected before you plan anything around it. Open it in a PDF viewer and try to select a paragraph of text. If you cannot select text at all the document is a scan, which is a different problem covered below. If you can select it but the file refuses to upload, that is the protection flag.

Pick the tool from the pile your question is in. Diagram question, manual under 100 pages: Claude. Diagram question, manual over 100 pages: Gemini Notebook with the figure added separately as an image. Text question of any length: any of them, and Gemini Notebook if you want it to stay available tomorrow without spending an upload.

Ask with the page numbers your viewer shows you. Claude’s documentation is explicit about this — when referring to specific pages, use the PDF page numbers as shown in your PDF viewer, not the page numbers printed on the document itself. In a manual the two never agree, because the cover, the legal notices and the table of contents all come before printed page 1. Asking about “page 94” when your viewer says 98 sends the model to the wrong table, and it will answer from the wrong table.

Ask narrowly, and ask for the location. “What does the amber symbol with the exclamation mark inside a circle mean, and which page is it described on?” beats “what is this warning light”. The page reference is the part that matters: it turns the answer into something you can check in four seconds instead of trusting.

When the diagram is the answer

Scanned manuals and figure-heavy pages are where all of this gets thin, and the honest answer is that you work around the tools rather than through them.

The reliable move is to stop feeding the whole document and feed the one page instead. Crop the figure, or photograph the page off the screen if cropping is awkward, and add the image as its own source — Gemini Notebook accepts image files directly as sources, which makes this a two-minute job. It also sidesteps the page limits entirely: the image is its own source, so the length of the document it came from stops mattering.

One trap worth knowing: Claude’s documentation notes that for non-PDF documents it extracts text only, and “if they contain embedded images, Claude won’t be able to read or interpret them”. A manual someone has helpfully converted into a Word file has had its diagrams removed from the model’s view entirely, even though you can see them perfectly well on your screen.

If the manual is genuinely confidential — a service document under NDA, or equipment schematics you would rather not upload anywhere — that is the case for running the whole thing locally instead. We have covered when a local model is actually worth the setup and how to get one running with Ollama; feeding it documents is a job for an interface like Open WebUI rather than the chat box.

What to check by hand, every time

Anything that can hurt you or cost you goes back to the page. Torque figures, electrical ratings, warning light meanings, fluid types, dosage on anything medical, load limits. Not because these tools are unusually bad at those questions, but because the cost of being wrong is asymmetric and verification takes seconds once you have a page number. This is the entire reason for asking where the answer came from.

Part numbers deserve their own warning. They are the single most fabricated thing in this whole workflow: strings of digits with a plausible shape, sitting in a context where a wrong answer looks exactly like a right one. Check a part number against the parts diagram before you order anything.

And a limitation worth stating plainly, because nobody else seems to: there is no published evaluation of how accurately any of these tools answer questions across a range of real manuals. Every number in this article is a documented capability limit from the vendor’s own documentation — how many pages, how many megabytes, how many uploads. None of it is a measurement of whether the answers are right. The verification step is not optional politeness. It is the only quality control that exists.

That said, the base case is strong. A manual is a well-structured document written to be referenced rather than read, full of exactly the short, locatable facts these tools are built to retrieve, and sitting on your drive precisely because nobody wants to read it. The answer to the amber light was on page 94 the whole time. The tools are good at finding page 94. Just make sure the one you picked can see the picture on it.