A trained document already works. Someone opens the link, types a question, gets a cited answer. Nothing else is strictly required.
But before they type anything they meet a title, a category, maybe a welcome message, maybe a disclaimer they have to accept, and, invisibly, an access rule that decided whether they got in at all. That is the part you configure, and it is the part that decides whether a first-time visitor asks a question or closes the tab.
Two things to find before anything else
Open a document, choose Edit, and the editor's tabs group into Content, Configuration and Learning. There are more of them than anyone wants to memorize, so two shortcuts matter more than the tour itself.
Search settings… in the header, or Cmd+K, searches every setting by name and shows its current value inline. If you can remember what a thing is called but not where it lives, this is the answer, and it saves the usual game of opening four tabs in sequence.
The Checklist tab is better still, and it is where a freshly trained document should start. It sorts every setting into Essential, Recommended and Optional, shows what each is currently set to or a placeholder where it is empty, and filters down to just the rows needing attention. Click a row and it jumps you to the right tab. Most products make you discover an empty intro message by noticing a bare-looking page. This one hands you the list.
One thing to know about saving, because the editor does it two ways. Toggles and dropdowns save the moment you change them, with a Saving… spinner and then a green Saved pill. Text fields, along with the access level and passcode, wait for Apply or Save & Close. Try to close with unsaved changes and it offers to save or discard rather than losing them.
Identity, and the address people will see
The Overview tab holds the document's name and its URL. The URL is the short identifier at the end of your share link, something like kdigo-ckd-2024, and it is worth setting deliberately rather than accepting whatever was derived from the filename, because it is the part people will paste into emails. If your organization has a custom domain, the copy button hands you the branded version.
Translations is the quietly good one here. Language does not translate anything by itself, it swaps the reader-facing labels between English and French. But if you have genuine French and English versions of a document, linking them under Translations lets readers switch between them from inside the chat, without hunting for a second link that somebody forgot to send them.
What greets a reader
The Messages tab is where the pre-conversation experience gets built, and it has four distinct jobs.
Introduction is the welcome text in the chat sidebar. It takes basic HTML with a live preview. Beside it sits Regenerate abstract & keywords, which asks the AI for a fresh summary, sample questions and keyword cloud, then shows you the result before you commit to it. The keywords are sampled from the whole document rather than just its opening, which is why they are usually better than the ones you would have written from memory.
Composer prompt is a small card sitting above the chat input, shown before the first message. Give it a label like "Getting started" and a line of guidance, and you have solved the blank-cursor problem for people who arrived without context. You choose whether readers can dismiss it, whether it disappears once chatting begins, and how long a dismissal sticks: about a year, or just the current session.
Explainer is a short note telling readers that this assistant answers only from its training material and will say when it does not know. It is off by default and shows once per session. Leave the text empty for the built-in wording. It is describing the behavior covered in how grounded answers work, and on a document where accuracy matters, it sets expectations before the first question rather than after a disappointing one.
Disclaimer goes further. Switch it on and readers must accept before they can use the document at all, with I agree and I decline. There is also a medication warning override, so a clinical document can show or suppress that warning independently of your organization's default. If both explainer and disclaimer are on, the explainer comes first.
The tabs that change what the chat can do
Images manages the cover and the document's image library, including images extracted during training. Its Gallery Tool sub-tab picks and orders which of them appear in the reader's lightbox. Empty the gallery and the gallery toggle switches itself off, which is a small kindness: no reader ever opens a gallery containing nothing. Videos works the same way for Vimeo links. See images and video.
Helpers adds optional questions above the chat so visitors can supply structured context, patient details for instance, that then travels with every message they send. Off by default. Turn it on and you get a field builder plus an option to show the questions as a step-by-step wizard the first time someone focuses the chat box. See helpers and guided context.
Abstract generates an AI overview of the whole document, in the background, with a cooldown between regenerations. Two details worth knowing: generating one does not show it to anyone until you enable the Abstract tool, and if you retrain the document afterwards, the tab flags that the abstract may now be out of date.
Who gets in, and who can find it
The Settings tab is the one with real consequences. Access level offers six choices, from Public through to Owner admins only, and the trade-offs between them are their own post: access levels explained.
Two more toggles decide where the document turns up. Show in owner library puts it in owner-scoped browse lists, and Show on landing page puts it on your organization's landing page. Both default to on, and a hidden document is still perfectly reachable by direct URL, so this is about tidiness rather than security.
Visibility & Search handles the rest. Discoverable in search lets a document appear in Docutrain's search results, but only if it is also Public, and a live status line tells you whether it currently will rather than leaving you to work it out. Suggest related documents and Include in related suggestions control the two directions of related-document suggestions when an assistant cannot answer, which is what powers the "this may be answered elsewhere" panel readers see after a refusal.
That covers everything a reader meets before typing. What happens afterwards, meaning tone, memory and which tools appear, is in tuning the chat experience.