common, server : add model type to GET /models

[angt]

Overview

Add a type field to GET /v1/models and GET /models.

The field gives the kind of the model:

• classifier: a native decision model. Use /v1/systemone for it.
• model: a normal model. Its metadata has no decision type.
• unknown: the metadata is missing or not valid.

Additional information

earendil-works/pi#10382

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

[ngxson]

UX-wise, I think having a single type is not very future-proof

In the past, we had the same story with models capable of doing infill/embeddings/reranking and combination of them. The verdict was to have an array of tags to indicate model's capability.

Maybe we can do the same here? Models capable of both clasification+chat are totally possible and high chance we will see some popping up soon

[ggerganov]

Does it make sense this to be input or output modality, similar to text, vision, audio etc?

[angt]

Something like that ?

{
  "architecture": {
    "input_modalities": ["text"],
    "output_modalities": ["decisions"]
  }
}

see openrouter: https://openrouter.ai/docs/guides/overview/models

[ggerganov]

This makes sense to me, but would defer to @ngxson for the final say.

[pcuenca]

openrouter skips decision models in the list unless explicitly asked for, not sure if we want to to do the same.

There's also a proposal to add capabilities to the OpenAI API, but no idea if it's getting accepted.

[ggerganov]

Given that OpenRouter does it like this, I think it's one more reason to follow that - better than adding our alternative at this point?

@mitsuhiko Would it work for you if the decision type is announced in the output modalities?

openrouter skips decision models in the list unless explicitly asked for, not sure if we want to to do the same.

Hm, don't see what is the reason to do that. Returning everything by default should be OK.

[pcuenca]

Yes, both make sense to me (use the output modalities, return everything).

[ngxson]

yeah I think the output_modalities would be the best

[angt]

Superseded by #29987