System Prompt Does Not Utilize dspy.Signature Docstring (__doc__)

[yanmxa]

The dspy.Signature class docstring (doc) is not dynamically parsed or integrated into the system prompt.

Steps to Reproduce:

1. Define a dspy.Signature subclass with a detailed docstring, e.g.,:

class KubeEngineer(dspy.Signature):
    """
    You are a Kubernetes engineer. Your task is to solve the given issue or perform the actions to interact with the Kubernetes cluster. 
    Use tools or kubectl commands to achieve the desired outcomes.
    
    Try to identify the following variables from the provided task or issue:
    - Cluster name
    - Resource type (e.g., Pod, Service, Deployment)
    - Resource name
    - Namespace
    """

    task: str = dspy.InputField(
        desc="The task or issue to address within the Kubernetes cluster."
    )
    answer: str = dspy.OutputField(
        desc="The final solution or output for the given task."
    )

2. Run the following code:

lm = dspy.LM(
    model="llama-3.1-70b-versatile",
    api_base="https://api.groq.com/openai/v1",
    api_key=os.getenv("GROQ_API_KEY"),
)
with dspy.context(lm=lm):
    react = dspy.ReAct(KubeEngineer, tools=[kube_executor_with_configs.kubectl_cmd])
    result = react(
        task="***"
    )
print(lm.history)

3. The system prompt in the output

{
  "prompt": None,
  "messages": [
    {
      "role": "system",
      "content": "Your input fields are:\n1. `task` (str): The task or issue to address within the Kubernetes cluster.\n2. `trajectory` (str)\n\nYour output fields are:\n1. `reasoning` (str)\n2. `answer` (str): The final solution or output for the given task.\n\nAll interactions will be structured in the following way, with the appropriate values filled in.\n\n[[ ## task ## ]]\n{task}\n\n[[ ## trajectory ## ]]\n{trajectory}\n\n[[ ## reasoning ## ]]\n{reasoning}\n\n[[ ## answer ## ]]\n{answer}\n\n[[ ## completed ## ]]\n\nIn adhering to this structure, your objective is: \n        Given the fields `task`, produce the fields `answer`."
    },
    {
      "role": "user",
      "content": "...."
    }
  ]
}

Current Behavior:

The system prompt does not include the docstring from the KubeEngineer class.

Expected Output:

{
  "prompt": None,
  "messages": [
    {
      "role": "system",
      "content": "You are a Kubernetes engineer. Your task is to solve the given issue or perform the actions to interact with the Kubernetes cluster. Use tools or kubectl commands to achieve the desired outcomes.\n\nTry to identify the following variables from the provided task or issue:\n- Cluster name\n- Resource type (e.g., Pod, Service, Deployment)\n- Resource name\n- Namespace ..."
    },
    {"role": "user", "content": "...."}
  ]
}

[okhat]

Hey @yanmxa ! Thank you so much for opening an issue & PR for this.

I think you're observing this in ReAct only, not in general. The issue can be addressed by editing predict/react.py

[okhat]

We need to pass instructions to this constructor:

dspy/dspy/predict/react.py

Line 67 in 698d0ac

dspy.Signature({**signature.input_fields, **signature.output_fields})

[yanmxa]

Thanks, @okhat! You got it! This issue occurs when using the ReAct module. Are there plans to address this? If so, I can draft a PR to resolve it.

[okhat]

Thank you, @yanmxa ! A PR would be welcome :D The change should be a one-liner but you may need to check that it does work.

For line 67, it should become:

dspy.Signature({**signature.input_fields, **signature.output_fields}, signature.instructions)

[yanmxa]

@okhat Thank you for the prompt response!