Retrain to improve accuracy
Turn the requests your model missed into training pairs and train a better model, on the Retrain tab.
Your users' real requests are the best training data a project can get. The Retrain tab lists them, lets you correct the calls, and turns the ones you choose into training pairs for another training run.
Retraining is a Pro feature. A project's first training run is free on every plan, and so is a retry when that first run fails.
The routine
Open the project's Retrain tab. It lists recent requests, newest first, each with its utterance and the calls the model returned.
Find the requests worth training on. Needs calls shows the requests that got no tool call, which are the best place to start.
Correct each call in its box. The calls are JSON: a list of {"name": …, "arguments": {…}}. A box shows Invalid JSON until the edit is valid.
Select the rows you corrected, and any rows the model got right that you want to reinforce, then choose Save N pairs.
Choose Train now (N new) to start a run on all the project's approved pairs, including the new ones.
When the run finishes, deploy it from Training jobs. A retrained model isn't deployed until you do. See Train and deploy.
A run started from Retrain uses the project's base model and settings from its first run.
Finding the right requests
| Control | What it shows |
|---|---|
| Search utterances… | Requests whose utterance contains the text. |
| All commands | Filters by the tool the model called. <unknown> shows calls to tools that don't exist in your project. |
| Needs calls | Requests the model produced no calls for. |
| Corrected | Requests a user fixed in the intent bar before running the call. |
| Collapse duplicates | Groups identical utterances into one row, so you can save or reject a group at once. |
| Show dismissed | Requests you rejected earlier. Choose Restore to bring them back. |
A row marked fallback was answered by the AI fallback, not by your model. Its calls are worth checking: if they're right, saving them teaches your model to answer that request itself.
Corrections from the intent bar
When a user changes the intent bar's chips before running a call, the request shows Corrected by a user, with the user's call beside the model's. Changed values are highlighted, and a tool the user switched away from is struck through.
The correction doesn't replace the model's call unless you choose Use this correction. Users can be wrong too, so check it first. See Build an intent bar.
Rejecting requests
Reject removes the selected rows from the list, and the X on a row removes that one. Use it for test traffic, nonsense and requests you've decided not to support. Rejected rows are hidden from History too, but still count in Analytics. Show dismissed lists them, and Restore brings them back.
Coming soon
Auto-score and AutoTrain, which would score and select rows automatically, are shown but not yet available.
Revised