Files and documents

Attach a PDF or a text file with the paperclip in the message box, then ask your question. The file is read when you attach it, not when you send, so you can see what was pulled out of it before committing to a question about it.

What you can attach

PDFs, and plain-text formats: .txt, .md, .csv, .tsv, .json, .log, .yml.

Not Word, Excel, or PowerPoint. Those are archives of XML rather than documents, and reading them properly is a different job from reading a PDF — exporting to PDF or pasting the text works today. Not images either: attach those directly instead, which is a separate button and a different path.

Two kinds of PDF

This is the part worth knowing, because it changes what you get.

A PDF that carries its own text — anything exported from a word processor, a browser, or almost any modern tool — has its text read directly. That is exact, it is fast, and it costs nothing beyond the text itself. Most PDFs are this kind.

A scanned PDF has no text in it at all. Every page is a photograph, and the only way to read one is to look at it. Those pages are handed to a model that can see, and you will see the chip say "read as images" rather than giving a character count.

Unizune AI works out which kind it has and takes the cheap path whenever it can. Running everything through a vision model would be simpler and worse: slower, more expensive, and less accurate on text that was already machine-readable.

Limits, and what happens at them

• 10MB per file. A PDF larger than that is almost always mostly images.
• 60 pages of text read from one PDF.
• About 40,000 characters per message across all attached documents — roughly

fifteen to twenty pages of dense text. This is sized against what the engine can actually take, not against what a PDF can contain.

• 3 documents per message.
• 4 pages of a scanned PDF, because each page is a whole image in the request.

When something is cut short you are told so, in the chip and in a notice. That is the point of the limits being visible: an answer about half a document is worse than an answer that says which half it read.

What happens to the file

Nothing is stored. The file is read when you attach it and then discarded — there is no copy on our side to fetch back, delete, or leak. What is kept is the extracted text, on the message, exactly like the words you typed.

One consequence worth knowing: the extracted text stays with the message you attached it to. Ask a follow-up question in the same conversation and it is still there. Start a new chat and you will need to attach the file again.

Getting a better answer

• Ask something specific. "What is the notice period?" beats "summarise this" by a

wide margin on a long document.

• Attach the relevant document, not all of them. Three documents share one budget,

so three long ones means less of each.

• Check the chip. If it says a document was truncated, the answer only covers what

was read — say which section you care about and attach that part instead.

• For a scan, fewer pages is better. Four pages of images is already a large

request; splitting a long scan and asking about each part works far better than attaching the whole thing.

Older documents in a long conversation

A document is much larger than a message, so replaying every one on every turn would fill the whole conversation with files you have stopped asking about. The most recent document is kept in full; earlier ones are remembered by name only.

If you go back to an earlier document after a long conversation, attach it again.