How to Chat With a PDF Without Uploading It Anywhere

Chatting with a PDF is genuinely useful. You drop in a fifty-page lease, ask when the break clause kicks in, and get an answer with a page reference instead of an afternoon of skim-reading.

The catch is that the documents worth interrogating this way — contracts, medical letters, financial statements, case files — are exactly the ones you should think twice about uploading to someone else's server. This guide shows how to do it with an AI model that runs inside your own browser, so the file never leaves your computer, and when the cloud tools are the better call anyway.

Why the usual tools want your file

Every mainstream chat-with-PDF service works the same way. Your document is uploaded, the text is extracted and split into chunks, each chunk is indexed on their infrastructure, and your questions are sent to a cloud model along with the relevant passages. That is a sound design and it produces good answers. It also means your document now exists on hardware you do not control, subject to a retention policy you probably have not read.

For a published research paper, who cares. For a settlement agreement or a patient record, that is a decision you may not have the standing to make on your own.

The local alternative, in three steps

Browsers can now run AI models directly on your graphics hardware, which means a web page can offer this without ever receiving your file. Our chat with PDF tool works this way:

Open the tool and pick your PDF. The file is read by the page itself, the same way a photo you attach to an email is read before you hit send. It is not transmitted.
Wait once for the model. The first visit downloads a compact language model into your browser cache. This takes a couple of minutes on a normal connection and does not repeat — later visits start immediately.
Ask your questions. The document is searched locally, the most relevant passages are fed to the model, and the answer comes back with the pages it relied on so you can expand and check them.

Proving it is actually local

You do not have to take our word for it, and you should not take anyone's word for it. Two tests settle it in under a minute:

The airplane-mode test. Load the tool and let the model finish downloading. Then turn off your wifi and ask a question. It answers. Nothing that requires a server can do that.
The network-tab test. Press F12, open the Network tab, and load your PDF. You will see the model files download, and you will not see your document go anywhere. There is no upload request because there is no upload.

These tests work on any tool making this claim, ours included. We wrote up the longer version in what happens when you upload a PDF.

Getting better answers out of a small model

A model that fits in a browser is smaller than a frontier cloud model, and it rewards precise questions. What works:

Ask specific questions, not open ones. "What is the notice period for termination?" beats "tell me about this contract." Retrieval finds a clause far more reliably than it finds a theme.
Use the document's vocabulary. If the lease says "lessee," ask about the lessee, not the tenant. The search step matches wording.
One thing at a time. Three questions asked separately beat one compound question.
Always open the cited pages. Every answer shows its sources for a reason. Small models can misread, and verification takes seconds when the passage is right there.
Run OCR on scans first. If your PDF is a photographed or scanned document, there is no text to search until it has been recognised. Our OCR tool adds a text layer locally, and then chat works normally.

When to use a cloud AI instead

Being straight about this matters more than winning the comparison. Reach for a frontier cloud model when:

The document is public or non-sensitive and you want the strongest possible analysis.
You need hard reasoning across the whole document — comparing several agreements, following a complex financial argument, catching subtle inconsistencies.
You are working with very long documents where a larger context window genuinely helps.
Your hardware is old or low on memory, where a local model will be slow.

The sensible pattern is to sort by document, not by tool. Public and complex goes to the cloud. Confidential stays local, where a good-enough answer you can verify is worth more than a brilliant one that required handing over the file. If you want the detail on what you are agreeing to with cloud tools, see AI and privacy.

Related things you can do locally

Chat is one of several jobs that no longer need an upload. If you want the whole document condensed rather than questioned, the AI summarizer does that on-device too. To pull invoice or receipt fields into a spreadsheet, use data extraction. And if you need to share a document with the personal details removed first, automatic redaction finds and strips them in your browser.

The useful shift here is that privacy stopped being a feature you have to trust and became a property of how the tool is built. A model running on your own machine cannot leak your document, because it was never given one to leak.