Feature request: Ability to train the Ultimate (AI Agents Advanced) bot model to improve answers that it gives

The feature being requested:

We are testing Ultimate bots (recently renamed to Zendesk ‘AI Agents Advanced’).  We are using the option Generative AI with all options selected (including ‘Activate zero-shot interface’ and ‘User zero-shot AI model’).

We find that sometimes it gives inaccurate or misleading answers.  We would like a way to train the Model to improve specific answers.

Background information and use case:

I realise that we can create ‘Use cases’ to push certain questions down a specific path towards documentation that give the right answer, but it would be good to be able to train the bot in circumstances where it has given a specific bad answer.

We can access ‘Conversation logs’, and in there we can 'Provide feedback'.  When we do that we can mark the answer as one of the following:

- Correct/well answered

- Incomplete answer

- Incorrect article used

- Made-up facts

- Incorrect language or vocabulary

- Other

It would be fantastic if the above feature actually trained the model to avoid the answer that was given.  However,  that is not ‘training the model’.  Instead it is described as  “This feedback is analyzed internally to improve UltimateGPT’s performance”.    Zendesk Support (in ticket 13343928) confirmed that this option is not used for training of the bot specifically, but more to 'help Zendesk's backend developers train the overall models'.

Currently there is no feature available to tell the bot that a specific answer is bad (ie. train the model).  Using ‘Use Cases’ to guide the bot for topics that lead to frequent bad answers is not really feasible, as there are many different scenarios where the answer is wrong, or slightly wrong.