How Mark reads your files: inside Saanora's document canvas
Attach a PDF, a Word file, a deck or a spreadsheet and Mark opens it beside the chat — and shows you exactly what it managed to read, so you know before you trust the answer rather than after.
You attach a forty-page report, ask about page 31, and get a confident answer about something else. Or you upload a slide deck and Mark misses the one diagram the whole thing was built around.
Neither is random. Both follow from what actually happens to a file between your machine and the model — and almost no AI product will tell you. Saanora now shows you, in the chat, on the same screen.
Files become text, and text has edges
Models read text. They do not open files. So every attachment goes through a conversion step first, and what survives that step is what Mark actually gets.
For a Word document, a spreadsheet or a presentation that means extraction: the words come through, most of the layout does not, and anything that exists only as a picture does not come through at all. A slide whose entire content is a screenshot arrives as an empty slide. PDFs and images are different — those can be read natively, as pages — but the general rule holds. What you see is not always what Mark sees.
There is also a ceiling on how much of one document reaches the model, and it is measured in characters rather than megabytes. A 5 MB PDF and a 50 MB PDF can deliver exactly the same amount of text, because most of what makes a PDF large is images.
“See what Mark read”
So Saanora stopped hiding the seam. Open a spreadsheet or a presentation in the canvas and there is a control marked “See what Mark read”.
It is not a preview of your document. It is a receipt. If a section is missing there, Mark could not see it either — and you know that before you rely on the answer instead of after you have already quoted it in an essay.
Most products treat this as an implementation detail. We think it is the single most useful thing you can know when an answer about a document looks off.
A canvas beside the chat, not on top of it
The canvas opens on the right and pushes the conversation across rather than covering it, so you can read the file and the answer about the file at the same time. It never covers the sidebar, even at full width — opening a document should not cost you your navigation.
While a file is being fetched and converted, you get a “preparing preview” state rather than a blank rectangle, because a blank rectangle is indistinguishable from a bug.
What renders in place:
- PDFs — embedded and readable, page by page.
- Word documents — converted and laid out as a page, in your browser. The bytes never leave your machine to be rendered.
- Plain text, Markdown and code — with syntax highlighting across the languages you would expect.
- Images — full size, with a lightbox.
- Everything else — an honest download panel. A browser cannot render an .xlsx, and pretending otherwise produces a broken frame instead of a clear next step.
View-only, on purpose
The canvas is a reader, not an editor. You cannot change a document inside it, and that is deliberate: an editor that silently diverges from the file Mark actually read would make every answer about it untrustworthy.
Download is always available on every type, because “look at this” and “keep this” are different needs and the second one should not require guessing.
The same canvas for what Mark makes
It works in the other direction too. When Mark builds you something — a Word document, a deck, a spreadsheet, a PDF — it opens in the same canvas, and downloads directly on Windows and Android rather than bouncing you out to a browser to fetch it.
Diagrams and 3D scenes get their own full-screen view with zoom and pan, for the ones that are too dense to read inline.
Getting more out of an upload
A few habits get more out of any assistant, Mark included:
- Ask about a named section. “What does the Methodology section say about sampling?” beats “summarise this”.
- Split very long documents. Two focused uploads beat one that gets truncated halfway.
- Export presentations to PDF when the layout carries meaning — extraction throws arrangement away.
- If the content is a picture of text — a scan, a photographed whiteboard — say so, and check the receipt.
- Put files a whole run of chats will need on a project instead of re-attaching them each time.