The claim on the box is that Docutrain answers only from your uploaded documents and cites its sources. That claim is worth something only if you can see the machinery. This post follows one question from the moment you press Enter to the buttons under the finished answer.
Watching an answer arrive
You type into the composer at the bottom of the chat page. The placeholder reads Ask a question, or Ask (Name) a question when the document has an assistant character. Enter sends; Shift + Enter adds a line break. Questions are capped at 1,500 characters, with a countdown appearing once you are within 200 characters of the limit, turning amber and then red as you close in.
After you send, your question appears under a Your Question label, followed by a thinking indicator — some documents show rotating tips while you wait. If retrieval takes more than a moment, Cancel and Retry buttons appear under it. Then the answer streams in word by word, and the send button becomes a square Stop Generating control that halts the answer but keeps the text produced so far. Occasionally the assistant pauses and shows "Taking a moment to think this through…" — it is working the question a second way behind the scenes, and needs nothing from you.
Answers can stop early, and a notice always explains which kind of early. Hitting the length limit suggests Succinct mode or a narrower question, with a Try again button; other interruptions offer Reload page; a dropped connection produces "Connection lost. Please check your network and try again." If the document itself fails to load, you get An error occurred with a Retry button instead of an empty chat you cannot use.
One thing you will not find is a model picker: the model is chosen by the platform and the document owner, and the dropdown in internal development builds never reaches readers. What you can choose is the response style, Chatty or Succinct. Each answer carries a badge naming the style that produced it, and tapping that badge offers Redo in Succinct or Redo in Chatty, which regenerates the answer in the other style in place rather than appending a new one. If the owner's custom prompt asks for long, detailed answers, your Succinct choice still wins.
The citations, and how to check them
As the assistant writes, it cites. Finished answers carry small Ref chips inside the text — "Ref · 03" and so on — and a list of the citation numbers is appended at the end.
Selecting a chip opens the source viewer, which is the part that matters. Each source appears as a card labeled Source text containing a verbatim excerpt from the document, with a note reminding you that formatting is condensed for retrieval and display. Chips above the excerpt show the reference number, the Page number in the original, and, in multi-document chats, the Document it came from. When one citation covers several sources, Previous reference and Next reference arrows move between them. Excerpts load on demand, so you will briefly see "Loading excerpt…". On a phone the viewer fills the screen edge to edge, which is the difference between reading a long excerpt and pinching at one.
Two caveats. Owners can turn the reference display off entirely, in which case answers appear without chips — a setting, not a failure. And these numbered citations are not the References tool, which is a bibliography the owner writes by hand; for the original file rather than an excerpt, use the Files tool. Both are covered in the tools that sit beside the chat.
When the document does not have the answer
This is the behavior worth testing before you trust anything else. If the assistant cannot find relevant information, it says so under a heading reading Not in this document, followed by:
I couldn't find that in my training. I only answer based on what's in these materials — I won't guess beyond that. Try asking about something covered here, or rephrase your question.
The refusal is not a dead end. If you opened the document from a collection, an extra line suggests the answer may be in one of the collection's other documents and points you at Chat with all docs in the collection. If the document belongs to an organization with others, the chat may search those and show an Elsewhere panel — "This may be answered in other documents:" — listing suggestions with their category, year, and page. Ask there opens that document's chat and asks your question again automatically.
Inside a collection, that panel is limited to the collection's own documents, and only ones you may open, so a suggestion never quietly walks you out of the collection you were given.
What else an answer can contain
Answers are not only text. Images from the document load with a shimmer placeholder and open full screen when tapped, with arrows to move between all the images in that answer; each caption sits in a footer band under the picture, like a figure in a published paper. Where the owner added a video gallery, a strong match to your question can bring a playable video card into the answer. See images and video.
When a question could go in more than one direction, the assistant answers what it can, then asks a short clarifying question followed by tappable option chips — "Which population would you like?", with options like "Adults" and "Special populations". Tapping a chip sends that choice immediately, and the original subject is kept in mind, so a two-word label is understood in context. The chips are suggestions, not the only routes.
The row of buttons under each answer
| Button | What it does |
|---|---|
| Export response | Opens a menu with Copy to Clipboard and an email-a-PDF option |
| Copy answer | Copies the answer text to your clipboard |
| Try Again | Re-asks the question and replaces this answer in place |
| Share | Creates a public link to the conversation; hidden in embeds |
| Thumbs up | Marks the answer helpful; select again to remove it |
| Thumbs down | Marks it unhelpful and opens a feedback form |
The thumbs-down form is titled What went wrong? and offers seven reasons you can combine: Factually inaccurate, Too short, Not helpful, Too slow, Off topic, Confusing, and Missing references. Submit Feedback records them, Skip feedback records the rating alone. Those reasons are not decoration — they land in the owner's conversation viewer as tags, the most direct available signal about which parts of a document are thin. See conversations and moderation.
Two more things can appear nearby. On medical documents a small hazard icon opens a flyout with the applicable notices: a Caution section headed Verify medication details before prescribing when the answer contains medication recommendations, and a Notice section headed Adverse event when your question looks like one — that second flag can arrive a few seconds after the answer finishes. Separately, a small circled "i" button near the bottom of the page on desktop opens the organization's legal links.
An assistant character, where one is configured, puts a name and avatar beside answers ("(Name) Says:"). That is presentation only. The answers underneath are still grounded in the document, and the citations are still the thing to check.