The first half of the document editor decides what a reader meets before they type. The second half decides what happens once they do.
Almost all of it is set per document, which is the useful part. Two assistants in the same workspace can behave nothing like each other: one terse and clinical, one chatty and branded, with different tools in the toolbar and different rules about what it will discuss.
The single setting worth your time
If you change one thing after training a document, make it the Custom prompt. It is the cheapest way to make an assistant sound like yours without touching a word of the content.
It takes up to 3,000 characters of plain instruction. Something like "Always respond in formal English. If a question is outside this document's scope, recommend the user consult their advisor." Tone, language, form of address, formatting preferences: all fair game.
What it cannot do is override the grounding. Your instructions layer on top of Docutrain's core rules, so citations, sourcing and the refusal behavior always win. You can change how an assistant sounds, not whether it tells the truth about what it knows.
That boundary is enforced rather than assumed. On save, the prompt goes through an automated safety review, and a rejected one comes back with the category and the reason so you can edit and retry rather than guess. Instructions that try to override grounding, invent facts, or direct abusive behavior are the ones that get turned away.
One more precedence rule, and it is the right one: a reader's Succinct choice beats your prompt. Ask for long, thorough answers all you like, and a visitor who wants the short version still gets the short version.
Tone and memory
Default chat tone sets where readers start. Chatty gives full-length answers with headings and context. Succinct gives brief, scannable replies, which suits quick lookups and phone screens. Readers can switch mid-conversation whatever you choose, so treat this as an opening position rather than a rule.
Memory is off by default and worth understanding before switching on. Turned on, it uses a signed-in reader's own earlier conversations with this same document as background context, so a follow-up question next week lands in the middle of a thread rather than at the beginning. It does nothing for anonymous visitors, and it never crosses between readers or between documents. One person's history is never another person's context.
It is a separate thing from Prior chats, which simply lets readers browse their own history themselves.
Chat attachments, where your organization allows them, let a reader attach a PDF or Word file as temporary context for their own conversation. Their file is not added to your document's training, which is the distinction that matters: readers can bring their own material to a question without permanently altering what everyone else's assistant knows.
The contact form that refuses to appear
Buried in the Engagement settings is one of the more opinionated behaviors in the product.
You can let readers email you from the chat. Recipients are entered as pills, the first being the To address and the rest CC'd, with the sender set as reply-to so you can just hit reply.
Then: every recipient address has to confirm it is willing to receive that mail. Each pill shows where it stands, from Send link through Link sent to a check mark, with a resend once the cooldown passes and a seven-day expiry on the link. Verification is remembered across your whole organization, so an address confirmed once is ready on every other document.
The consequence is the interesting part. If the To address is missing or unverified, the contact tool does not appear on the chat page at all. No reader ever types a message into a form that was never going to deliver it. Unverified CC addresses do not block anything and are quietly left off until they confirm.
It is a small piece of restraint, and it is the difference between a contact form and a shredder.
What sits in the toolbar
The UI Options tab decides which tools readers get, and it has a habit worth noticing: options stay disabled until they would actually work.
The image and video galleries cannot be switched on until you have configured gallery items. The Abstract button stays greyed out until an abstract exists. Quiz waits until questions have been generated. Recent questions needs at least two suitable ones before it displays anything. You cannot accidentally offer a reader a button that opens onto nothing.
Citations are on by default and append source references to answers. On documents assembled from several source files, or extended through added-content retraining, this can be turned off with an explanatory note, because page references stop being reliable when the pages came from four different originals. It is better to show no page number than a wrong one.
References is the other half of the sourcing story, and a different thing entirely: your own curated bibliography, one entry per line, with URLs and DOIs turned into links automatically, shown behind its own chip.
Most of these accept English and French label overrides, so Downloads can become "Handouts", the gallery "Slides", References "Sources". What readers do with the resulting toolbar is chat tools and the sidebar.
Giving an assistant a face
Character, available where the chat assistant feature is enabled for your organization, puts a name and avatar beside answers. The avatar can be uploaded or generated from a text description, and the description is kept so you can regenerate it later without rewriting it from scratch.
Set one here and it fully replaces your organization's character for this document, which is how a single workspace ends up with a formal assistant on the compliance manual and a friendlier one on the onboarding guide.
It is presentation only. The answers underneath are grounded exactly as before.
Where everything else lives
The Attachments tab manages downloadable files, and quietly includes a Download Activity card tracking who downloaded what, when, and from where, exportable as CSV. For anyone distributing controlled material, that log is often the reason to attach files here rather than emailing them.
The rest of the editor is covered elsewhere in this series: Embed & Share in embedding on your website, Telegram in Slack, Telegram and custom domains, Quiz in quizzes, Retrain, History and Logs in retraining and history, Conversations in conversations and moderation, and Intelligence in document intelligence.
One tab belongs to super admins only: Metadata, a read-only panel of timestamps, raw configuration and internal identifiers kept for support. If you never open it, nothing is wrong.