Training a document is the short part. What follows is the part you live with: fourteen assistants in a table, two of them paused, one that needs a different access level, and a colleague asking for the onboarding link while you are on a call.
The library is built for that afternoon rather than for the first one.
Three chips you can edit without opening anything
Most of the routine work never requires opening a document at all, because three of the table's columns are editable in place.
Status reads Live, Paused, or Quota, and clicking toggles it. Pausing is reversible and deletes nothing. A paused document even stays in any collection it belongs to, marked Paused and counted separately rather than silently dropped, so pausing something for a fortnight does not quietly unpick it from four collections you would then have to rebuild from memory.
Quota is worth recognizing on sight. It means the document was paused automatically because your organization has more live documents than the plan allows. Nothing was lost; something needs to give.
Access changes who can open the document without going near the editor. Public, Passcode, Registered, Owner, or Owner Admins Only, with public and passcode requiring a plan that includes them. See access levels explained.
Category opens a searchable picker with your categories, then the system defaults, and a + Custom… option to create one on the spot rather than abandoning what you were doing to go and set up a taxonomy first.
Together those three cover most of what you actually do to a document after it exists.
Finding things
Search matches title, slug or document ID, and Cmd+K jumps straight to the box. Filters cover status, access level and category, where choosing a parent category also matches its subcategories.
The useful detail: filter choices are stored in the page address. So a filtered view can be bookmarked, or pasted to a colleague who then sees exactly what you were looking at. "The three public onboarding documents" becomes a link rather than a set of instructions.
Row counts go up to 500 per page, and translation pairs are kept together in the ordering.
The row menu, and the honest one
The three-dot menu holds the expected things: view, configure, conversations, intelligence, delete.
Download PDF is the one worth pointing at. It appears only when the source file is still stored, and if the document was trained from more than one file it lists every version, the original first, then each retrain upload with its date and size. That is a small archive of everything you ever fed the assistant, which is more than most tools will tell you six months later.
Copying a link uses your custom domain when you have one. Copy the link of a non-public document and you get a reminder that visitors will still need to meet its access requirements, before you paste it into an email rather than after.
Combining documents
Tick some checkboxes and a bulk bar offers Combine, Create collection, and Delete.
Combine merges several trained documents into one new one, which is how a set of chapters or related guidelines becomes a single assistant instead of five. It needs at least two documents from one organization and a free slot.
The window is unusually honest about what you are getting. The originals are untouched. The new document is a snapshot, so it will not update when the sources are retrained, and it cannot be retrained itself, because it has no source file. To refresh it you combine again and delete the old one.
The copy runs from your browser, so the progress bar comes with a plain warning not to close the tab. Each copied passage is prefixed with its source document's title, which means citations in the combined assistant still tell a reader which original they came from. That is the detail that makes a combined document trustworthy rather than a blended mush.
If a combine never finishes because the tab was closed mid-copy, the half-built document sits in the library with a red, unclickable Incomplete chip. Delete it and start again. It tells you rather than pretending to be a working assistant.
Deleting
Deletion is permanent and takes the processed content and embeddings with it. Both confirmation windows make you type DELETE.
Bulk deletes run one at a time with a running counter, and if anything fails the window stays open and names each failure rather than closing on a cheerful summary that was not true.
If you only want something out of sight, pause it. That is what pause is for, and it frees no plan slot, which is usually the reason people reach for delete instead.
Documents created by publishing a form are the exception: they carry a Form badge, have no retrain tab, and deleting either half deletes both plus every response.
The dashboard around it
Owner admins land on a Home page rather than straight into the table. Tiles for documents, collections and forms; quieter shortcuts to users, intelligence, and your plan usage.
Two things there are better than they need to be. The users tile grows an amber N pending chip when people are waiting for approval, which matters because join requests are not emailed to anyone. And a count that cannot be read is left out entirely rather than displayed as a misleading zero.
A Public landing page card gives you your organization's address with copy and open buttons, and a line above it saying which address you are actually copying: your custom domain, or a Docutrain one because DNS verification has not completed yet. No wondering why the link you sent looks wrong.
Recent activity lists the five most recently changed documents and collections. Retraining does not add a second row for the same document, so the feed stays a list of things rather than a log of events.
Categories, covers, abstracts
Categories power the library filter, your readers' browse views and the inline chips. There are system categories maintained centrally, and custom ones your organization creates, both two levels deep. Deleting a parent deletes its subcategories, which the confirmation spells out before you agree to it.
Covers are managed on the editor's Images tab: upload, generate from a keyword prompt, reuse an image already in the document, or fall back to your organization's default. See images and video.
The Abstract tab writes an overview from the processed content in the document's own language. A French document gets a French abstract and French keywords, which is the sort of thing that is invisible when it works and infuriating when it does not. Regeneration has a 24-hour cooldown, and retraining afterwards flags the abstract as stale.
Everything inside the editor itself is the document settings tour.