What is a model draft?

Every change you make to a classification or extraction model is saved as a draft. The draft is yours to work on for as long as you need: incoming messages keep being processed by the version you published last, and nothing you change reaches live traffic until you publish it.

This article explains what a draft is and how to recognize one. To put a draft live, see the "Publish a model draft" article.

Why drafts exist

Editing a model used to mean editing the model that was running. Renaming a label or rewriting a description took effect immediately, which made it risky to work on a model during business hours.

With drafts, the model has two sides:

• The published version is the one that processes live messages. It only changes when you publish.
• The draft is your working copy. It holds every edit you have not published yet.

You can leave a draft sitting for days, hand it to a colleague, or throw it away. The published model keeps running the whole time.

How to tell a model has a draft

Open the model and look at the title. A blue Draft badge next to the model name means there are unpublished changes.

The Publish button carries a colored status dot that tells you the same thing at a glance:

• Amber - there are unpublished changes.
• Green - everything is published and the draft is clean.
• Gray - the model has never been published.

A model with no unpublished changes shows no badge, and the button sits at green and reads Published.

What lands in the draft

These changes go into the draft and wait for you to publish:

• The model name and description.
• Adding, editing, and deleting labels (classification) or entities (extraction).
• Label and entity descriptions.
• Attachment processing settings.
• Importing entities from a CSV file on an extraction model.

A few settings are not part of the draft and take effect right away, because they do not change what the model predicts:

• Whether the model is tracked in analytics, and which inboxes it is tracked for.
• Your example set. Confirming or correcting an example updates your examples immediately.

Accuracy while you have a draft

Once a model has both a published version and a draft, the header shows the draft's accuracy first and the published figure in brackets, for example 90% (Published: 82%). Selecting the figure opens the accuracy panel, where Published model and Draft model appear as two cards you can switch between.

The card you select changes the breakdown below it, so you can see exactly which labels or entities the draft improves before you commit to it. In the draft view, counts that moved are struck through and the new number is shown next to them.

Leaving a model with a draft open

Navigating away from a model that has unpublished changes brings up a reminder. Nothing is lost either way: the draft is saved on the server, so closing the tab or reloading the page keeps it.

• Leave as draft continues where you were going and keeps the draft.
• Stay here cancels and returns you to the model.
• Don't show this again for any model turns the reminder off for your account. It applies from the next time you leave, not this one.

Which models have drafts

Drafts apply to classification and extraction models. Matching models are configured and saved directly and do not use the draft workflow.

Related articles

• To put a draft live, see the "Publish a model draft" article.
• To throw a draft away, see the "Discard a model draft" article.
• To see what changed and go back to an earlier version, see the "View a model's version history" article.
• What are models on Tekst?