feat(gemini): support structured outputs

[mattmcdev]

Created this PR before realising this one existed.

The main difference is this PR updates the structured output to match the schema format Google defines. It then passes this directly using response_schema rather than appending a message.

To me this seems the preferred way of doing it:

• The schema matching the defined format is going to help the model know how to use your schema.
• Passing the schema via response_schema means you get validation (and relevant errors) on your schema

[kinsta]

Preview deployments for prism ⚡️

Status	Branch preview	Commit preview
✅ Ready	Visit preview	Visit preview

Commit: e1162ff023a34f722dd08f3730f464a5012946c4

Deployment ID: d871fe03-fdf3-4c20-8f31-f68cd128707a

Static site name: prism-97nz9

[ChrisB-TL]

I'm not yet familiar with the Gemini API but if it accepts a proper schema that would seem the better route.

Do you know if there limitations at all, which would make the Anthropic adaption approach better?

We use the Anthropic adaption a lot, and whilst it works most of the time it isn't completely reliable. It goes off piste around 5% of the time.

[mattmcdev]

I'm not yet familiar with the Gemini API but if it accepts a proper schema that would seem the better route.

Do you know if there limitations at all, which would make the Anthropic adaption approach better?

We use the Anthropic adaption a lot, and whilst it works most of the time it isn't completely reliable. It goes off piste around 5% of the time.

Google's docs seem to encourage using the schema via model configuration:

Important: This more formal method for declaring the JSON schema gives you more precise control than relying just on text in the prompt.

The only potential (sort of) downside is your call will fail if the schema doesn't pass validation which wouldn't happen with the Anthropic approach - personally I'd prefer to know that my schema was invalid and correct it.

[ChrisB-TL]

Yea, I see that it would fail at the request level as an advantage - you don't get billed for the tokens for a broken request/response.

[mattmcdev]

Yea, I see that it would fail at the request level as an advantage - you don't get billed for the tokens for a broken request/response.

Yep, and additionally I see the 'passing schemas as a message' type approach as a kind of hacky workaround for something that models seem to slowly be adopting more officially. So if there is an official implementation, probably better to work with it.

[pushpak1300]

I think we should support both the implementation as both are supported by gemini at the moment. #173 similar like openai structred support.

src/Providers/OpenAI/Handlers/Structured.php

[benbjurstrom]

Seems like response_schema is the way to go here. What's the use case for supporting both implementations?

@mattmcdonald-uk appreciate the pull request. Just pulled it into a project and it's working great with gemini-2.0-flash.