Retraining a document, and reading its history

The handbook got a new edition. The guideline was revised. Someone finally wrote the FAQ that answers the question your assistant keeps fumbling.

Retraining updates the content behind an existing assistant without disturbing anything around it. The document keeps its link, its title, its access rules, its galleries and every other setting you configured. Nobody has to be told a new URL.

Retrain when the material is the same subject in a newer or fuller form. Upload a new document when it is genuinely a different subject, which keeps each assistant focused and gives the new material its own link and its own analytics.

Add or Replace is the whole decision

Add is the default because it is reversible and Replace is not. Reach for Replace when the new file genuinely supersedes the old one and you do not want the assistant blending outdated and current material into one confident answer.

Two safety behaviors are worth knowing before you press anything.

Replace does not take your assistant offline. The previous content keeps answering questions the entire time the new content is processing, and the switchover happens in a single step at the end. A Replace that fails partway leaves the old content live rather than leaving you with a broken document and an audience.

Your abstract and keywords are protected. If the automatic regeneration fails during an Add retrain, the document keeps the abstract, keywords and welcome message it already had. A failure in the nice-to-have does not destroy the thing you wrote by hand.

Replace also clears the image gallery and rebuilds it from the new content, slide renders included, so you are never left with figures from a version of the document that no longer exists.

Running one

Begin Retraining on the Retrain tab opens a window with the same input tabs as a fresh upload, plus a mode selector that shows the current chunk count, so you can see what you are adding to or replacing.

The file tab takes the usual formats and asks the same Document or Slide deck question for PDFs. The audio tab transcribes on Business and above. The text tab has a note under the box spelling out exactly what is about to happen: either "All existing chunks will be deleted and replaced", or "New chunks will be added to the existing N chunks". No ambiguity at the moment it matters.

One consequence of pasting text: it has no page numbers, so source page references are switched off afterwards. You can re-enable them in the document settings once you know what you are doing.

Two kinds of document cannot be retrained at all, and both say why rather than just hiding the button. A chat document paired with a form has no source file, since its knowledge comes from submissions. A document created by combining others is a snapshot, and refreshing it means combining the sources again.

While it runs

You can leave the page. Processing continues, and the card reports the real stage rather than a static "please wait".

Two steps take noticeably longer than the rest, and knowing which is which saves a lot of unnecessary worry. Scanned PDFs are read by more than one service, with the Logs tab recording which reader produced the text. Complex diagrams get an extra pass, because arrows and branching structure do not survive ordinary text extraction.

Several documents can retrain at once, but two runs on the same document are refused outright. If a run was interrupted by a server restart, its lock releases by itself after about five minutes.

Most stalls resolve themselves, since Docutrain checks for interrupted runs at startup and every fifteen minutes. A run with no activity for over five minutes earns a Stuck badge and a Force Retry button, and a failed one offers a retry. You are rarely stuck waiting on someone else to unblock it.

What changes for people already chatting

Answers reflect the new content as soon as processing finishes. Conversation history and analytics are untouched.

Recent questions get handled with more care than you might expect. If a visitor clicks a question that was answered before the retrain, Docutrain re-asks it against the current content rather than replaying the cached answer, because that old answer might cite pages that no longer exist. A stale answer with a confident citation is worse than a slow one.

Three things want your attention afterwards. The Abstract tab flags itself as stale, so regenerate it. Quizzes are not regenerated automatically, since the existing questions were written from the previous content (quizzes). And images, videos and covers stay as they are, although a slide-deck retrain can add newly extracted slides.

The history is the undo button

The History tab records every training event as a timeline. Initial is the original. Add appended content and can be undone. Replace rebuilt from scratch and cannot.

Each row shows the operation and status, a download link for that session's actual file, the input type, how many chunks it created, the size, the duration, and the date. Every file ever used to train the document is kept, each retrain upload stored separately rather than overwriting the last, so you can always go back and look at what you actually fed it in March.

Each row also takes a note of up to 2,000 characters. That field is the cheapest insurance available against a colleague asking, six months later, why the guideline document changed and who decided it should.

Undo appears on a retrain when the session was an Add, it completed, and its chunks still exist. The confirmation states how many chunks will be permanently deleted before you commit. Only that session's content goes.

One limitation to plan around: a later Replace wipes earlier Add sessions, and after that they can no longer be undone. If you are layering supplementary material onto a document and might want to peel a layer back off, keep using Add.

Confirming it worked

When it finishes, ask the assistant two questions: one that only the new content can answer, and one that the removed content used to answer.

Thirty seconds, and it tells you more than any status badge.