Should there be maxTemperature, analog to maxTopK?

[tomayac]

Currently, as per #17 (comment), the temperature of the Prompt API is "by definition a number between 0 and 1," but, in practice, different models tend to have different ranges. GPT-4o models use a range between [0, 2], and here's an example for various Gemini models:

Range for gemini-1.5-flash: 0.0 - 2.0 (default: 1.0)
Range for gemini-1.5-pro: 0.0 - 2.0 (default: 1.0)
Range for gemini-1.0-pro-vision: 0.0 - 1.0 (default: 0.4)
Range for gemini-1.0-pro-002: 0.0 - 2.0 (default: 1.0)
Range for gemini-1.0-pro-001: 0.0 - 1.0 (default: 0.9)

To reduce developer confusion for developers coming from such cloud-based APIs to the Prompt API, would it make sense to add a new optional maxTemperature attribute to the AIAssistantCapabilities interface?

readonly attribute unsigned long? defaultTopK;
readonly attribute unsigned long? maxTopK;
readonly attribute float? defaultTemperature;
readonly attribute float? maxTemperature; // new

Knowing the maximum allowed temperature would also help with building a slider-based UI for the model configuration, as is common, for example, with OpenAI's Playground or Hugging Face's Gemini Pro Playground.

[domenic]

Yep. I erroneously thought temperature was always from 0.0-1.0, but it turns out you can plug other numbers into $\Pr(v_k) = \frac{e^{l_k / \tau}}{\sum_i e^{l_i / \tau}}$, with interesting results. Adding shortly.