Asking questions about your documents
Before you start
A folder open with documents in it, and an AI set up — a MindMax account or your own API key. The first time a folder is opened, PDFs, books and Office files are read in the background; until a file has been read in, answers say how many are not ready yet.
Asking about one document
With a document open, ask about it in the assistant panel beside it (⌘L / Ctrl+L). The open
document is offered to the conversation without pasting. Short documents are read in their
entirety rather than sampled, so nothing is left out. Longer ones are navigated by their chapter
structure, and the relevant sections are read in full. Each answer lists what it read.
Asking about a folder
Type @ and pick a folder, and its documents become the subject of that question. This is what
you want for "what do these papers disagree about" — a question no single document answers.
With nothing mentioned, the assistant searches your whole knowledge base when it needs to.
Asking about part of a page
Select a passage and ask about just that. Or drag a box around a chart, a table or a figure and ask about the region — the image goes with the question, so you do not have to describe what you are looking at.
A boxed region carries more than the picture: the verbatim text underneath it, so numbers and formulas are taken from the text rather than read off the image, and the page or chapter it came from, so clicking the thumbnail later reopens the document at that spot. The region stays with the conversation, so a follow-up about the same figure does not start from nothing.
Narrowing what it reads
Documents you @ are must-read for the whole conversation, instead of depending on being found
again each turn. The line above the input says what the scope is:
- Default — the documents you added come first; the knowledge base is searched when needed.
- These N only — only the documents you added. No knowledge base, no web. If they do not hold the answer, it says so instead of reaching further.
- Open it up returns to the default.
To keep a file out of every conversation, hide it from AI — see choosing what the AI can read.
When it looks things up online
If your folder does not hold the answer, the assistant can search the web for the gap, and the pages it actually cited are listed as numbered references at the end of the answer. When an answer rests on a figure that dates, it re-checks it in the background and appends a correction if needed.
Checking the answer
Claims carry citation marks. Hover one to read the quoted sentence with the text around it, as it stands in the document now; click it to open the document at that passage, highlighted.
Check the two claims your work will actually rest on. That is 90 seconds, and it converts an answer you cannot evaluate into one you can. Checking everything is not the point; checking nothing is how errors get through.
Quotes are matched against the source text before the answer is finished. A quote that is close but not exact is snapped back to the document's own wording; one that cannot be matched at all is flagged at the end of the answer rather than quietly presented as real. That catches invented quotes; it does not catch a real quote supporting a claim less well than it appears to.
When answers are disappointing
| Symptom | Usually |
|---|---|
| "I could not find that" on something you know is there | The document is a scan with no text layer — recognise it |
| Answers about the wrong document | Too broad a folder — @ the right documents, or use These N only |
| Confident answer, no useful quotes | The question has no anchor in the text; rephrase towards something the documents actually discuss |
| "N files are not ready yet" | They are still being read in — this finishes on its own |
Next
- Searching your folder — when you want the passages, not an answer
- Handing the assistant a long job
- Reading and annotating
- Writing in the editor