The handbook got a new edition. The guideline was revised. Someone finally wrote the FAQ that answers the questions your assistant keeps fumbling. Retraining updates the content behind an existing assistant without changing anything else — the document keeps its link, title, access rules, galleries and every other setting.
Retrain when the material is the same subject in a newer or fuller form. Upload a brand-new document when it is genuinely a different subject, which keeps each assistant focused and gives the new material its own link and analytics.
Add or Replace
Every retrain runs in one of two modes, and choosing between them is most of the job.
| Add to existing data (default) | Replace all data | |
|---|---|---|
| Existing content | Kept; new content appended after it | Discarded once the new content is ready |
| Abstract and keywords | Regenerated from everything, old and new | Regenerated from the new content only |
| Requirements | The document must already have content | None |
| Can it be undone? | Yes, from the Training History | No; the previous content is gone for good |
| Confirmation | Starts immediately | A warning window asks you to confirm |
| Typical use | Supplementary material | A new version of the document |
Two safety behaviors are worth knowing before you press anything. Replace does not take your assistant offline: the previous content keeps answering questions until the new content is fully processed, then the switchover happens in one step, and a Replace that fails partway leaves the previous content live. And your abstract and keywords are protected — if the automatic regeneration fails during an Add retrain, the document keeps its previous abstract, keywords and welcome message.
Add is the default for a reason: it is reversible and Replace is not. Use Replace when the new file genuinely supersedes the old one and you do not want the assistant mixing outdated and current material.
Running one
Open the document in the editor — its title in the Documents Library, or Configure from its actions menu — and select Begin Retraining on the Retrain tab.
The Retrain Document window opens with three input tabs — PDF / DOC, Text and Audio — and a Retrain mode selector showing the current chunk count, so you know what you are adding to or replacing. Provide your content, then Start Retraining. If you chose Replace, a Confirm Replace All Data window warns that all existing chunks will be permanently deleted; Yes, Replace All Data proceeds. Replace also clears the image gallery and rebuilds it from the new content, slide renders included, so you never end up with stale images from the previous version.
The file tab accepts PDF, DOCX, PPTX, XLSX and MD; PDF page counts are checked against your plan, and you are asked What type of PDF is this? — Document or Slide deck, the latter extracting slide images so the assistant can show them. PowerPoint files still upload with text extraction only. The Text tab takes 10 characters and 5 words as a minimum, and a note under the box spells out what will happen: "All existing chunks will be deleted and replaced", or "New chunks will be added to the existing N chunks". Pasted text is stored with the training record rather than as a file; because it has no page numbers, source page references are turned off afterwards, and you re-enable them in the document settings. The Audio tab takes MP3, WAV, M4A, OGG, FLAC and AAC on Business and higher plans.
Two kinds of document cannot be retrained at all. A chat document paired with a form has no Retrain tab, because its knowledge comes from form submissions; a Form chat document note points you to the Forms tab. A document created with the Combine bulk action is a snapshot with no source file, so a Combined document note replaces its Retrain section: to refresh it, combine the sources again and delete the old snapshot.
While it processes
A retrain repeats the training pipeline on your new content: stored, extracted or transcribed, chunked, indexed, saved. Add mode appends the new chunks and regenerates the abstract and keywords from the combined content; Replace mode prepares everything in the background and cuts over in a single step. Each retrain upload is kept as its own timestamped file, so earlier versions are never overwritten.
You can leave the page; processing continues. If the system is at capacity the status shows "Queued for processing...", though the card reports the real stage rather than a static label. Several documents can be in flight at once, but two runs on the same document are refused: "This document is already being processed." If a run was interrupted by a server restart, its lock releases automatically after about five minutes.
Two steps take noticeably longer than the rest. Scanned PDFs are read by more than one service, and the Logs tab records which reader produced the text. Complex diagrams get an extra pass, because arrows and branching do not survive ordinary text extraction.
Most stalls resolve themselves, since Docutrain checks for interrupted runs at server start and every fifteen minutes. A run with no activity for over five minutes gets a Stuck badge and a Force Retry button, Cancel stops a run in flight, and a failed run offers Retry retraining.
What changes for people already chatting
Answers reflect the new content as soon as processing completes, and conversation history and analytics are untouched. Recent questions are handled carefully: if a visitor selects a question that was answered before the retrain, Docutrain re-asks it against the current content instead of replaying a cached answer that might cite material no longer present.
Three things want your attention afterwards. The Abstract tab flags itself "Stale — document has been retrained since this abstract was generated", so regenerate it. Quizzes are not regenerated automatically, since the existing questions were written from the previous content (see quizzes). Images, videos and cover images stay as they are, although a Slide deck retrain can add newly extracted slides.
Training History
The editor's History tab records every training event as a timeline, oldest first. Initial is the original document the assistant was built from. Add appended content and can be undone. Replace rebuilt the assistant from scratch and cannot.
Each row shows the operation, a status (Completed, Failed, Started, or Deleted for an undone session), the Source — a download link for that session's file, or View pasted text — the input type, the chunks created (with a "+N" suffix on Add retrains showing how many already existed), the file size, the duration with token usage, and the date. Every row can carry a note of up to 2,000 characters, which is the cheapest available insurance against a colleague asking in March why the guideline document changed in September.
Undo appears on a retrain only when the session was an Add, it completed successfully, and the chunks it created still exist — a later Replace wipes earlier Add sessions, after which they can no longer be undone. The Undo Training Session? window states how many chunks will be permanently deleted; confirm with Delete Chunks. Only that session's chunks go, and the undo cannot itself be reversed.
The neighbouring Logs tab records every stage of every training, retraining and quiz generation run — rarely needed day to day, worth its weight when a run fails. And every file ever used to train the document is kept: the Download PDF action in the Documents Library lists the Original upload followed by each retrain upload in order.
Owner admins can retrain any document in their organization. Super admins can retrain any document on the platform, and additionally see a Bypass training limits toggle that lifts the plan-based limits for that run.
When you are done, ask the assistant a few questions that only the new content can answer, and one or two that the removed content used to answer. That is the fastest confirmation the retrain did what you meant.