Guide

How to Chat With a PDF Offline on iPhone, iPad, and Mac

Most "chat with PDF" tools work the same way: you upload your file to a server, the server does the reading, and answers come back over the network. That is fine for a recipe. It is a problem for a lease, a medical record, or a client contract — and it is useless on a plane. This guide walks through doing the whole job on the device itself: what hardware you need, the actual steps, how to judge whether the answers deserve your trust, and where on-device AI honestly falls short.

"Offline" is really two different needs

People searching for an offline PDF chat usually mean one of two things, and it is worth separating them because they have different stakes.

The privacy need: the document must not leave your device. Uploading a confidential file to a cloud AI creates a copy on someone else's infrastructure, subject to their retention policies, their subprocessors, and their security. Even a trustworthy provider is a second party you now have to reason about — the confidential documents guide goes deeper on exactly what you are agreeing to when you upload.

The connectivity need: you simply do not have a network. A flight, a job site, an archive basement, a secure facility where phones stay in airplane mode. Cloud tools fail here not because of policy but because of physics.

On-device AI answers both needs with the same move: the model runs where the file already is, so nothing needs to travel. If you want the background on why this only recently became possible on phones and laptops, see what on-device AI can actually do.

What you need before you start

The honest gate is hardware. Running a language model locally takes a capable neural engine and enough memory, and Apple only ships its on-device foundation models to devices that can handle them. To follow this guide you need:

• iOS, iPadOS, or macOS 26 or later, and
• Apple Intelligence-capable hardware: an iPhone 15 Pro or later, or an iPad or Mac with an M1 chip or later.

That floor is real, not marketing segmentation — a base iPhone 15 or a pre-M1 iPad cannot run these models, and no app setting changes that. If your device does not qualify, skip ahead to the limits section rather than installing anything; the rest of this walkthrough will not work for you.

Beyond that, all you need is a PDF. Text-based PDFs work directly; scanned PDFs work too, because on-device OCR can read the page images — though scan quality matters, as covered below.

Step by step: import the PDF, ask a question, tap the citation

The walkthrough uses OpenIntelligence, which is my own app, so weigh that as you would any author recommending their own work. It is free to download, and it has the same requirements just listed: iOS/iPadOS/macOS 26 or later on an iPhone 15 Pro or later, or an M1-or-later iPad or Mac. It needs no account and no API key, because there is no server to sign in to. The retrieval engine it runs is open source, and described on the home page if you want to see how the machinery works.

1. Import the PDF. On iPhone or iPad, use the share sheet from Files, Mail, Safari, or wherever the PDF lives; on Mac, drag the file into the app. The app parses the document, runs OCR on any scanned pages, and builds a local search index. This indexing happens once per document, entirely on the device, and takes longer for big or heavily scanned files — let it finish before asking questions.
2. Ask a question in plain language. Specific questions get better answers than vague ones. "What is the termination notice period?" will beat "tell me about this contract." You are not searching for keywords — you can ask about meaning, and the retrieval step finds the relevant passages even when they use different words than you did.
3. Tap a citation. Every claim in the answer carries a tappable citation. Tap one and you land on the exact passage in the original document that supports the claim. This step is the entire point — do not skip it, especially for anything you plan to act on.

How to tell the answers are trustworthy

This applies to any document AI, cloud or local. A trustworthy tool exhibits two behaviors, and you can test for both in about a minute:

Citations that resolve. An answer with no citations is an assertion; an answer with citations you cannot inspect is decoration. What you want is a citation that takes you to the actual sentence in the actual document — so the tool's claim and the source sit side by side and you can see whether they match. Spot-check a few. If a citation lands on a passage that does not say what the answer claimed, stop trusting that tool for that job.

Abstention. Ask a question you already know the document cannot answer — ask a lease about warranty coverage, or a physics paper about its authors' favorite food. A trustworthy tool says the document does not contain that information. A tool that produces a confident, fluent answer anyway is telling you how it will behave on the questions you cannot check.

OpenIntelligence is built around these two behaviors: it verifies drafted claims against the retrieved passages before showing them, and when the document does not support an answer, it declines and says so rather than filling the silence. That refusal can be mildly annoying in the moment. It is also the property that makes the tappable citations mean something.

Proving nothing was uploaded: the airplane-mode test

A privacy claim on a marketing page — including this one — is just a claim. The nice thing about on-device processing is that you do not have to take anyone's word for it:

1. Turn on Airplane Mode, and turn Wi-Fi off too.
2. Import a new PDF — one the app has never seen. This matters: testing with an already-indexed document only proves answering works offline, while a fresh import proves the entire pipeline does.
3. Ask your questions and tap the citations.

If everything works with the radios off, there is no cloud in the loop — not for indexing, not for retrieval, not for answering. A cloud-dependent tool fails this test at step two or three, immediately and unambiguously. This is also consistent with what the App Store privacy label for OpenIntelligence says — "Data Not Collected" — but the airplane-mode test is the stronger evidence, because it is one you ran yourself.

Beyond one PDF

Real questions rarely live in a single file. The same on-device pipeline extends to a library: import a folder of PDFs alongside Office and iWork documents, spreadsheets, presentations, notes, scans, code, and even audio or video recordings, and ask one question across all of them. The answer cites whichever documents contributed, so you can still trace every claim to its source. The multi-PDF search guide covers that workflow, including how to organize collections so retrieval stays sharp.

Honest limits — and what to do if your hardware doesn't qualify

A guide that only tells you what works is an advertisement. Here is the rest.

The hardware floor excludes a lot of devices. If you have an iPhone 15, an iPhone 14, an SE, or a pre-M1 iPad, this workflow — and my app — cannot serve you, and I would rather say that plainly than let you install into a wall. Your realistic options: run it on a qualifying Mac or iPad instead of your phone, or, if you must use a cloud tool, choose one deliberately with the retention and redaction practices in the confidential documents guide in mind.

On-device models are smaller than frontier cloud models. For grounded question-answering over your own documents — where every claim is checked against a passage — that tradeoff works in your favor. For sprawling open-ended synthesis across hundreds of pages, or long creative rewriting, a large cloud model will produce more fluent output. The on-device answer to a question the documents cannot support is an abstention, not a longer essay.

Scan quality is a real ceiling. OCR on a clean scan is reliable; OCR on a skewed, low-contrast photocopy of a photocopy is not, and no answer built on misread text can be better than the misreading. If a citation lands on garbled text, re-scan the source document properly.

One thing that is coming, stated carefully: support for Apple's Private Cloud Compute is built into the app but not live today. It arrives with iOS and macOS 27, expected in September 2026, and when it does it will be consent-gated, used only for final answer synthesis, and will show you the exact payload before anything is sent. Today, every route runs on-device — which is why the airplane-mode test above passes.

That is the whole picture. If your hardware qualifies and your questions live in your documents, you can have the convenience of asking in plain language without the file ever leaving your hands — and you never have to take that on faith, because you can switch the radios off and watch it keep working.