[Bug] JSONAdapter's parse method regex pattern bug

[hufeng]

What happened?

What happened?

Use dspy.ChainOfThought to execute task signature, as shown in the following example

class Issue(BaseModel):
    issue_type: str
    severity_level: str
    problem_code_snippet: str


class FindIssue(dspy.Signature):
    code = dspy.InputField(description="check code")
    issue_list: list[Issue] = dspy.OutputField(description="issue list")


class FindIssueModule(dspy.Module):
    def __init__(self):
        super().__init__()
        self.find_issue = dspy.ChainOfThought(FindIssue)

    def forward(self, code: str):
        return self.find_issue(code=code)

fim = FindIssueModule()
fim(code="...code...")

llm output

'{\n  "reasoning": "think step by step",\n  "issue_list": [\n    {\n  "issue_type": "some type",\n      "severity_level": "fatal",\n  "problem_code_snippet": "if (user) {"\n    }\n  ]\n}'

error message

Adapter JSONAdapter failed to parse the LM response.

{
  "issue_type": "some type",
  "severity_level": "fatal",
  "problem_code_snippet": "if (user) {"
    }
  ]
}

Expected to find output fields in the LM response: [reasoning, issue_list]

Problem diagnosis

I checked the code and output and found that the problem lies in the greedy matching behavior of the regular expression and the handling of nested curly braces. When the text contains curly braces within backticks (`), these curly braces are incorrectly treated as part of the JSON structure, interfering with the regular expression matching.

The parsing error is in https://github.com/stanfordnlp/dspy/blob/main/dspy/adapters/json_adapter.py#L150.

Steps to reproduce

The parsing error is in https://github.com/stanfordnlp/dspy/blob/main/dspy/adapters/json_adapter.py#L150.

test code

text = '{\n  "reasoning": "think step by step",\n  "issue_list": [\n    {\n  "issue_type": "some type",\n      "severity_level": "fatal",\n  "problem_code_snippet": "if (user) {"\n    }\n  ]\n}'


pattern = r"\{(?:[^{}]|(?R))*\}"
match = regex.search(pattern, text, regex.DOTALL)
if match:
    text = match.group(0)

print(text)

output:

{
  "issue_type": "some type",
      "severity_level": "fatal",
  "problem_code_snippet": "if (user) {"
    }
  ]
}

DSPy version

3.0.3

[hufeng]

Is it possible to simply maximize the algorithm to find the first { and the last } ，Currently the test works

def parse(self, signature: type[Signature], completion: str):
        # Find the first { and the last } in the completion
        first_brace_index = completion.find("{")
        last_brace_index = completion.rfind("}")

        if first_brace_index == -1 or last_brace_index == -1 or first_brace_index > last_brace_index:
            raise AdapterParseError(
                adapter_name="JSONAdapter",
                signature=signature,  # type: ignore
                lm_response=completion,
                message="Invalid JSON format: braces not found or mismatched.",
            )

        completion = completion[first_brace_index : last_brace_index + 1]
        fields = json_repair.loads(completion)
        .....