Python: Feat/Add DEVELOPER role for openai o1

[ymuichiro]

Motivation and Context

Currently, the OpenAI O1 model introduces a new role "developer". However, passing the conventional "system" role results in an error. To address this issue, the following changes are proposed.

Description

1. Add a Method to the ChatHistory Class

   • Implement a method to handle role conversion or substitution logic for compatibility with the O1 model.

2. Expand the AuthorRole Enum

   • Add "developer" as a new value to the AuthorRole enum.

3. Improve Developer Experience (UX)

   • If the "system" role is mistakenly passed, the logic should internally convert it to "developer" for better UX.
   • However, since all models other than O1 still use "system", simply adding support for the "developer" role seems to be the most optimal solution at this point.

Background

• The O1 model no longer supports the "system" role, causing errors when it is passed.
• The introduction of the "developer" role must be integrated in a way that preserves compatibility with other models.

Usage

import asyncio

from semantic_kernel.connectors.ai.function_choice_behavior import FunctionChoiceBehavior
from semantic_kernel.connectors.ai.open_ai.prompt_execution_settings.azure_chat_prompt_execution_settings import (
    AzureChatPromptExecutionSettings,
)
from semantic_kernel.connectors.ai.open_ai.services.azure_chat_completion import AzureChatCompletion
from semantic_kernel.contents.chat_history import ChatHistory
from semantic_kernel.kernel import Kernel

async def main():
    kernel = Kernel()
    service_id = "azure-openai"
    kernel.add_service(
        service=AzureChatCompletion(
            service_id=service_id,
            api_key=*********,
            endpoint=*********,
            deployment_name="o1",
            api_version="2024-12-01-preview",
        )
    )

    settings = kernel.get_prompt_execution_settings_from_service_id(service_id=service_id)

    if isinstance(settings, AzureChatPromptExecutionSettings):
        settings.function_choice_behavior = FunctionChoiceBehavior.Auto(auto_invoke=True)

    service = kernel.get_service(service_id=service_id)

    if not isinstance(service, AzureChatCompletion):
        raise Exception("Invalid Value")

    history = ChatHistory()
    history.add_developer_message("You are a helpful assistant.")
    history.add_user_message("hello")

    result = await service.get_chat_message_contents(chat_history=history, settings=settings, kernel=kernel)

    if not result:
        raise Exception("result is None")

    print(result[0].content)

if __name__ == "__main__":
    asyncio.run(main())

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄

[moonbox3]

Thanks for having a look at this, @ymuichiro. I think it’d be good to make sure other o1 api features are in as well. I saw they have a reasoning_effort param, too. Are there other features we’d want to get in to make sure the o1 experience works well?

[ymuichiro]

@moonbox3

Thank you for checking this so quickly. Regarding o1, here are the changes I’m aware of:

※ It might be worth considering providing OpenAIChatReasoningBase as a separate service from OpenAIChatCompletionBase in the future. However, since it seems that the functionality is being gradually integrated, I believe the appropriate place to incorporate these changes for now would be in OpenAIChatPromptExecutionSettings.

o1: 2024-12-17
API: 2024-09-01-preview

1. In o1, the system role has been deprecated, and a new developer role has been introduced.
2. In o1, max_tokens has been deprecated and replaced with max_completion_tokens.
   • max_completion_tokens: int
   • The value represents the sum of reasoning tokens and completion tokens, with a recommendation to specify 25,000 or more.
3. A new property, reasoning_effort, has been added in o1.
   • reasoning_effort: enum, low, medium, high
   • Higher values indicate that the model will spend more time processing the request.
4. A new usage metric, reasoning_tokens, has been introduced in o1.
   • This represents the number of tokens used during reasoning.

{
  "completion_tokens": 1843,
  "prompt_tokens": 20,
  "total_tokens": 1863,
  "completion_tokens_details": {
    "audio_tokens": null,
    "reasoning_tokens": 448
  },
  "prompt_tokens_details": {
    "audio_tokens": null,
    "cached_tokens": 0
  }
}

Currently Unsupported

• Parallel tool invocation
• temperature, top_p, presence_penalty, frequency_penalty, logprobs, top_logprobs, logit_bias

It would be nice if these were applied to settings, usage, and author_role.