How do I set Temperature to ChatCompletionAgent

[piffy76]

It seems ChatCompletionAgent takes PromptExecutionSettings as the execution settings which does not include Temperature, MaxToken, PresencePenalty, etc that are in OpenAIPromptExecutionSettings. I tried using OpenAIPromptExecutionSettings instead, but it seems the extra properties are being ignored. Is this by design?

[crickman]

I would not expect these to be ignored. I'll investigate and reply with my findings.

[piffy76]

Thank you. It also seems to ignore the system prompt if I use OpenAIPromptExecutionSettings. A dilemma is, if it respected the system prompt, wouldn't it be a conflict/redundancy to an Agent's Instructions property?

[crickman]

For ChatCompletionAgents the the agent instructions are translated as the system prompt (so yes, redundant...actually equivalent). I'm surprised to hear that your experience varies with regards to the PromptExecutionSettings. I'd expect that regardless of any PromptExecutionSettings, the presence of an additional system prompt could result in confilct or amibguity.

[piffy76]

To be clear, I had the impression that all the extra properties in OpenAIPromptExecutionSettings were being ignored because of the fact that the system prompt in the setting was clearly being ignored, but I did not examine the rest of the properties, such as Temperature, PresencePenalty, etc. Do you suggest that only the system prompt is ignored because it is a redundancy/overridden by the Instructions, but the rest of the properties will be respected? If so, that'd be understandable.

[crickman]

Yes, I believe so. Will do some verification on this end to make sure.

[crickman]

Confirmed that OpenAIPromptExecutionSettings and respected by ChatCompletionAgent (iterated on MaxTokens, TopP and Temperature.

You can confirm this by setting Temperature to a value greater than 2 or MaxTokens very low (i.e. 10). Either of these should result in an error (which demonstrates they are included in the model request).

[piffy76]

Thank you!