reasoning-update-phi-2507 #1875
Conversation
| assistant_message = Message( | ||
| role=response_message.role or "assistant", | ||
| content=response_message.content, | ||
| reasoning_content=response_message.reasoning_content |
There was a problem hiding this comment.
Can we make this an official field on Message?
There was a problem hiding this comment.
Because then this logic can be part of the main code and we don't have to overwrite here?
There was a problem hiding this comment.
reasoning_content is limited to just DeepSeek provider with iddeepseek-reasoner. As the scope is very limited, I just made this change for DeepSeek class
More here: https://api-docs.deepseek.com/guides/reasoning_model
There was a problem hiding this comment.
agree with dirk this should be an official field on Message
libs/agno/agno/reasoning/agent.py
Outdated
| def get_deepseek_reasoning_agent( | ||
| reasoning_model: Model, | ||
| monitoring: bool = False, | ||
| ) -> Optional["Agent"]: # type: ignore # noqa: F821 | ||
| from agno.agent import Agent | ||
|
|
||
| return Agent( | ||
| model=reasoning_model, | ||
| description="You are a meticulous and thoughtful assistant that solves a problem by thinking through it step-by-step.", | ||
| monitoring=monitoring, | ||
| ) |
There was a problem hiding this comment.
What makes this particular to deepseek? The description?
There was a problem hiding this comment.
Yes, I plan on refining the description and it does not need the extra params that we are providing to the default agent.
libs/agno/agno/reasoning/agent.py
Outdated
| @@ -15,6 +15,27 @@ def get_reasoning_agent( | |||
| tools: Optional[List[Union[Toolkit, Callable, Function, Dict]]] = None, | |||
| structured_outputs: bool = False, | |||
There was a problem hiding this comment.
I think we should specify specific structured output for the reasoning agent. And not assume it should be passed on from the base agent.
Tools as well, you might want specific tools for it. Also you might have like 50 tools, now it gets sent to your reasoning agent as well.
There was a problem hiding this comment.
I intentionally did not change the logic for our current reasoning. I think that fits in another PR
There was a problem hiding this comment.
we should 100% specify it but sadly not all models support it, so we need to lean on the user to set structured_outputs=True
then the reasoning agent uses structured_outputs=True
agree the tools thing is a mess, but currently we want the reasoning to be able to run the same tools so it can to the work using CoT and then give it to the final agent. lets have a chat on call about this
libs/agno/agno/reasoning/helpers.py
Outdated
| for msg in reasoning_agent_response.messages: | ||
| if hasattr(msg, "reasoning_content"): | ||
| extracted_reasoning = msg.reasoning_content | ||
| break |
There was a problem hiding this comment.
Couldn't there be multiple messages?
There was a problem hiding this comment.
One reasoning_content block with the reasoning and one content block with the actual content. Since we disable streaming for reasoning, this works
Nevermind I see that is a next step |
libs/agno/agno/agent/agent.py
Outdated
| if reasoning_model.__class__.__name__ == "DeepSeek" and reasoning_model.id == "deepseek-reasoner": | ||
| from agno.reasoning.helpers import get_deepseek_reasoning | ||
|
|
||
| # Ensure the reasoning model and agent do not show tool calls | ||
| reasoning_agent.show_tool_calls = False | ||
| reasoning_agent.model.show_tool_calls = False # type: ignore | ||
|
|
||
| step_count = 1 | ||
| next_action = NextAction.CONTINUE | ||
| logger.debug("==== Starting Reasoning ====") | ||
| while next_action == NextAction.CONTINUE and step_count < self.reasoning_max_steps: | ||
| step_count += 1 | ||
| logger.debug(f"==== Step {step_count} ====") | ||
| try: | ||
| # Run the reasoning agent | ||
| reasoning_agent_response: RunResponse = await reasoning_agent.arun( | ||
| messages=run_messages.get_input_messages() | ||
| ) | ||
| if reasoning_agent_response.content is None or reasoning_agent_response.messages is None: | ||
| logger.warning("Reasoning error. Reasoning response is empty, continuing regular session...") | ||
| break | ||
| reasoning: Optional[Message] = get_deepseek_reasoning( |
There was a problem hiding this comment.
testing this, but not particularly happy about his :/
Changes:
Next steps:
reasoning_functionparameter that allows users to pass a callable function. This function will be executed, and its output will be added as a message to the Agent. Alternatively, a new parameter can be introduced to accept a function whose result is directly added as a message to the Agent. This approach makes the Agent more flexible, enabling additional use cases beyond reasoning.print_response