QUESTION: It seems that LiteLLM does not handle the reasoning_content of the model.

[BreezeTwinesJade]

When I was connecting to the model via LiteLLM, I noticed that it doesn't seem to support outputting the model's reasoning_content. Here's the content of my code:

llm = LiteLlm(
    model=os.getenv("BASE_MODEL"),
    api_base=os.getenv("BASE_URL"),
    api_key=os.getenv("API_KEY"),
)

base_planner = BuiltInPlanner(thinking_config=ThinkingConfig(include_thoughts=True))

agent = Agent(
    name="simple_agent",
    model=llm,
    instruction="你是一个乐于助人的中文助手",
    description="回答用户的python编程问题",
    planner=base_planner
)

session_service = InMemorySessionService()

session = session_service.create_session_sync(
    app_name="adk_project",
    user_id="1",
    session_id="1"
)

runner = Runner(
    app_name="adk_project",
    agent=agent,
    session_service=session_service,
)

if __name__ == '__main__':
    content = types.Content(role="user", parts=[
        types.Part(text="我想要请求一个外部服务，通过http，你有什么好的建议？以'经过以下思考：'作为回答的开头")])
    print("开始调用模型")

    thinking = False
    for event in runner.run(user_id="1", session_id="1", new_message=content,
                            run_config=RunConfig(streaming_mode=StreamingMode.SSE)):
        print(event)

I saw the following code in the source code parsing. It seems that LiteLLM does not wrap the reasoning_content output by the model. Is there a problem with my usage or is it indeed not supported?

[stonyme]

qwen and deepseek not suport outputting reasoning_content,Even though Litellm supports Deepseek output reasoning_content。

[vipwp]

代码里都没有处理reasoning_content的逻辑

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: The BuiltInPlanner is designed for models with native thinking capabilities like Gemini. For other models accessed via LiteLLM, you should use the PlanReActPlanner to see the model's reasoning process.

---

Hello! Thanks for your question.

The behavior you're observing is expected. The BuiltInPlanner is specifically designed to work with models that have a native, built-in "thinking" or "reasoning" feature, such as Google's Gemini. It relies on the model providing its reasoning in a specific, separate part of the response that ADK can then surface as reasoning_content.

When you use a model through LiteLLM, it is likely that the underlying model does not support this specific native thinking feature. For these cases, ADK provides the PlanReActPlanner.

The PlanReActPlanner is a more model-agnostic solution. It works by instructing the model to follow a specific output format that includes its reasoning, planning, and actions directly in the main response. This allows you to see the thought process of a wider range of models.

To fix this, you should switch your planner from BuiltInPlanner to PlanReActPlanner. Here is how you can modify your code:

from adk.agents.planners import PlanReActPlanner

# ... your existing code ...

# Replace BuiltInPlanner with PlanReActPlanner
# base_planner = BuiltInPlanner(thinking_config=ThinkingConfig(include_thoughts=True))
base_planner = PlanReActPlanner(thinking_config=ThinkingConfig(include_thoughts=True))

agent = Agent(
    name="simple_agent",
    model=llm,
    instruction="你是一个乐于助不的中文助手",
    description="回答用户的python编程问题",
    planner=base_planner
)

# ... rest of your code ...

By making this change, you should be able to see the model's reasoning and thought process in the events generated by the runner.

---

This response was generated by an automated assistant. Please note that the information provided may not be exhaustive.