You can chat with a PDF on your phone, offline and without uploading it, by using an AI app that reads documents on the device. You attach the file, ask a question, and the app finds the relevant passages in the document and has a local language model answer from them. The PDF never leaves your phone, which matters for contracts, medical paperwork, tax forms and anything else you wouldn’t paste into a cloud chatbot.
It works best on text-based PDFs and specific questions. Scanned documents and “summarize all 300 pages” requests are where it struggles.
How does chatting with a PDF work on a phone? #
A phone-sized language model can only hold a limited amount of text in mind at once, its context window. A long PDF won’t fit, and even a short one would slow the model down. So document chat works in two steps:
- Find the right passages. The app extracts the text from the PDF, breaks it into chunks, and searches those chunks for the parts that match your question. This is the same idea behind “semantic search,” where passages are matched by meaning rather than exact words.
- Answer from those passages. Only the matching chunks go to the model, along with your question. The model writes an answer based on them.
The technique is called retrieval-augmented generation, or RAG. Its big advantage is that the model answers from your document instead of from memory, which cuts down on made-up answers. Its weakness is that if the search step misses the right passage, the model never sees it.
How to chat with a PDF offline, step by step #
Personal LLM has this built in. You attach a PDF, text or Markdown file, the text is extracted and searched on the phone, and the model answers from the matching passages and tells you which ones it used. Nothing is uploaded.
- Install the app and download a model while you’re on Wi-Fi. Qwen 3.5 4B (2.74 GB) is the recommended first download. Models with larger context windows help with documents; Ministral 3 3B is described as quick on documents and screenshots.
- Save the PDF to your phone. Files app on iPhone, Files or Downloads on Android.
- Start a new chat and attach the file.
- Ask a specific question. “What’s the cancellation fee?” works better than “tell me about this.”
- Check the passages it cites. They show you where the answer came from, so you can confirm it.
Once the model is downloaded, all of this works in airplane mode.
What questions work best? #
Specific questions with answers that live in one or two places in the document:
- “What’s the notice period for ending this lease?”
- “Does this policy cover water damage? Quote the relevant part.”
- “What dosage does this leaflet say for adults?”
- “List every deadline mentioned, with dates.”
- “What does ‘force majeure’ mean in section 12?”
- “Which fees apply if I pay late?”
Questions that need the whole document at once are harder, because the app is picking passages rather than reading everything. For “summarize this report,” it’s often better to ask section by section: “Summarize the introduction,” then “Summarize the findings.” Our guide to summarizing text offline has more on this.
What are the limits of offline PDF chat? #
Scanned PDFs #
A scanned PDF is often just pictures of pages with no text layer. There’s no text to extract, so document chat can’t read it. You can check by trying to select a word in a PDF viewer: if you can’t highlight text, it’s an image.
Workaround: take clear photos of the key pages and ask a vision-capable model about them instead. See asking AI about a photo offline.
Tables and multi-column layouts #
Text extraction can scramble tables, footnotes and two-column layouts. If an answer about a table looks off, check the table yourself.
Missed passages #
If your question uses different words than the document, the search step may pick the wrong chunks. Rephrase using the document’s own terms, or ask about a specific section.
Small models misreading #
Even with the right passages, a small model can misread a clause or mix up two numbers. That’s why the cited passages matter: read them before relying on an answer. Why AI makes things up covers this in detail.
Phone memory #
Very large documents take longer to process on a phone, and a bigger context size uses more RAM. On an older phone, keep documents smaller or split them.
Cloud vs offline PDF chat #
| ChatGPT, Gemini, Claude | Offline on your phone | |
|---|---|---|
| Where the PDF goes | Uploaded to the company’s servers | Stays on your phone |
| Handles long documents | Better, with larger models | Works by picking passages |
| Scanned PDFs | Often handled with cloud vision | Needs a text layer, or use photos |
| Answer quality | Stronger models | Smaller models, check citations |
| Works offline | No | Yes |
| Cost | Free tiers have limits | Free models, one-time download |
For a lease, a medical report or a bank statement, the privacy difference is the main reason to go offline. Our guide on sharing financial info with AI explains what’s at stake with cloud uploads.
Tips for better answers #
- Ask the model to quote. “Quote the sentence that says this” makes it easier to check.
- Use the Precise preset for factual questions about documents. It lowers the randomness in answers.
- Switch to a bigger model for tricky documents. In Personal LLM you can move a hard question to Qwen 3.5 9B from the chat header, if your phone has the RAM.
- One document per chat. Mixing files makes it harder to tell which one an answer came from.
- Start a new chat for a new document. Old context from a previous file can confuse the answer.
Frequently asked questions #
Can I chat with a PDF without uploading it? #
Yes. Use an app that processes documents on your phone, such as Personal LLM. The text is extracted and searched on the device, and a local model answers from the matching passages, so the file never leaves your phone.
Does offline PDF chat work with scanned documents? #
Not if the PDF has no text layer. Document chat needs text it can extract. For scanned pages, photograph them and ask a vision-capable model about the images instead.
How long a PDF can I use? #
Longer documents work because the app searches for relevant passages instead of reading everything at once. But very long files take longer to process on a phone, and questions that need the whole document are harder. Split huge files if you can.
Is offline PDF chat accurate? #
It’s more accurate than asking a model from memory, because the answer is based on your document. It’s not perfect: the search can miss passages, and small models can misread them. Check the passages it cites before relying on an answer.
Which file types can I ask about? #
In Personal LLM, PDF, plain text and Markdown files. Other apps vary. Word documents can usually be exported to PDF first.