Fix structured output dispatch and settings

[JonasHelming]

fixed #14810

What it does

• Move call to structured output variant before handling non streaming requests in openAI language model, otherwise, it is never used
• Introduce a parameter to model descriptions whether an opneAI model supports structured output
• Add default values in the frontend contribution where other defaults are specified
• Make the new parameter configurable for custom openAI models

How to test

Use the terminal agent with 4o and o1-preview and verify with a debugger that structured output is used only for 4o

Follow-ups

Breaking changes

• This PR introduces breaking changes and requires careful review. If yes, the breaking changes section in the changelog has been updated.

Attribution

Review checklist

• As an author, I have thoroughly tested my changes and carefully followed the review guidelines

Reminder for reviewers

• As a reviewer, I agree to behave in accordance with the review guidelines

[planger]

Thank you! Seems to work fine, except for a typo with gpt-4 (see inline).
I tested this with the Terminal agent and setting different OpenAI models to be used for this agent.

[planger]

By turning the whitelist into a blacklist, accidentally 'gpt-4o-2024-05-13' is now considered to support structured output, which it infact does not. Also I believe 'o1-mini' does not support structured output (https://platform.openai.com/docs/models#o1, just o1, but I might misinterpret the docs, haven't tested it).

I think expecting structured output, while there is no support for it currently yields an error in Theia AI at the moment. So I wonder whether going with blacklisting is a good idea. People can always use a custom model to work around that (in both directions) though, so I don't have hard feelings either way.

[planger]

Also gtp-4 -> gpt-4

[JonasHelming]

Good points!!

• typo is fixed
• I added "gpt-4o-2024-05-13" and "o1-mini"

I decided for a blacklist because it seems that all newer models are supporting structured output (with the exception of early o1 stream), so I believe this is less maintenance in the future. For the blacklisted ones, I tried to cover everything that might still be used and that we had previously in our default list. I now added 'gpt-4o-2024-05-13" even though I hope nobody is using this anymore.

In any case, if a model is configured wrongly, there will be an error and you can use a custom configuration. This will actually at the moment only hit users using an agent that uses structured output (the terminal is the only in the Theia IDE) AND manage to select an OpenAI LLM that does not support this, which I find unlikely. Tool builders have this under control anyways.

[planger]

Thanks, looks good to me! 👍