[Bug] program doesn't know how to use tools (ReAct)

[chrico-bu-uab]

What happened?

(See "Steps to reproduce" for code.)

When I compile/save my program, I see stuff like this:

        "trajectory": "[[ ## thought_0 ## ]]\nI need to determine the margin status of the breast cancer pathology report.\n\n[[ ## tool_name_0 ## ]]\nRegex\n\n[[ ## tool_args_0 ## ]]\n{}\n\n[[ ## observation_0 ## ]]\nFailed to execute: Regex.__call__() missing 1 required positional argument: 'report'\n\n[[ ## thought_1 ## ]]\nI need to determine the margin status of the breast cancer pathology report.\n\n[[ ## tool_name_1 ## ]]\nRegex\n\n[[ ## tool_args_1 ## ]]\n{}\n\n[[ ## observation_1 ## ]]\nFailed to execute: Regex.__call__() missing 1 required positional argument: 'report'\n\n[[ ## thought_2 ## ]]\nI need to determine the margin status of the breast cancer pathology report.\n\n[[ ## tool_name_2 ## ]]\nRegex\n\n[[ ## tool_args_2 ## ]]\n{}\n\n[[ ## observation_2 ## ]]\nFailed to execute: Regex.__call__() missing 1 required positional argument: 'report'\n\n[[ ## thought_3 ## ]]\nI need to determine the margin status of the breast cancer pathology report.\n\n[[ ## tool_name_3 ## ]]\nRegex\n\n[[ ## tool_args_3 ## ]]\n{}\n\n[[ ## observation_3 ## ]]\nFailed to execute: Regex.__call__() missing 1 required positional argument: 'report'\n\n[[ ## thought_4 ## ]]\nI need to determine the margin status of the breast cancer pathology report.\n\n[[ ## tool_name_4 ## ]]\nRegex\n\n[[ ## tool_args_4 ## ]]\n{}\n\n[[ ## observation_4 ## ]]\nFailed to execute: Regex.__call__() missing 1 required positional argument: 'report'",

Not sure if I am doing something wrong or if there is a bug. The documentation is very unclear about the args parameter in particular.

Thanks!

Steps to reproduce

I can't provide the data, but below is the tool function.

class Regex:
    """Simple tool that uses regular expressions to extract margin status."""

    name = "Regex"
    desc = "extracts margin status using regular expressions"
    args = {'report': {'type': 'string', 'description': 'text from a pathology report'}}

    def __init__(self):
        self.pattern = re.compile(r"positive margin|margin positive|negative margin|margin negative")

    def __call__(self, report: str) -> str:
        """Extract margin status from the report."""
        match = "|".join(self.pattern.findall(report.lower()))
        print("Match:", match)
        if match:
            return "positive" if "positive" in match else "negative"
        return "inconclusive"

DSPy version

2.6.9rc1

[okhat]

Hey @chrico-bu-uab ! What model are you using?

With 4o-mini, I run this and it works fine:

import re

pattern = re.compile(r"positive margin|margin positive|negative margin|margin negative")

def regex(report: str) -> str:
    """Extract margin status using regular expressions."""
    match = "|".join(pattern.findall(report.lower()))
    if match:
        return "positive" if "positive" in match else "negative"
    return "inconclusive"


r = dspy.ReAct('question -> answer', tools=[regex])
r(question="positive margin")

[chrico-bu-uab]

I tried SmallThinker:3b and DeepSeek-r1:8b, neither of which work on my data. The best model I have access to for this research is DeepSeek-r1:32b, which I haven’t tried yet.

[okhat]

I see. I'd avoid reasoning models for non-reasoning problems; they're known to, at least so far, be overfit to specific types of problems (math, coding) and less good at following arbitrary instructions.

You could use LLama-3.2-3B or 3.1-8B for instance. Or Qwen-2.5 models.