fix: guard against None input in trimAndLoadJson

[bongho]

Problem

When the evaluation LLM returns None (instead of a string), trimAndLoadJson crashes with:

AttributeError: 'NoneType' object has no attribute 'find'

This happens because the function calls input_string.find("{") without first checking whether input_string is None.

Related issue: #2554

Solution

• Changed the type hint of input_string from str to Optional[str]
• Added an explicit None guard at the top of the function that:
  • Raises a descriptive ValueError with the same message used for other JSON parse errors
  • Sets metric.error when a metric is provided (consistent with existing error handling)

Changes

 def trimAndLoadJson(
-    input_string: str,
+    input_string: Optional[str],
     metric: Optional[BaseMetric] = None,
 ) -> Any:
+    if input_string is None:
+        error_str = "Evaluation LLM outputted an invalid JSON. Please use a better evaluation model."
+        if metric is not None:
+            metric.error = error_str
+        raise ValueError(error_str)
+
     start = input_string.find("{")

Testing

The fix is minimal and follows the existing error-handling pattern already present in the function's except json.JSONDecodeError block.

[vercel]

@bongho is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.

[bongho]

Hi team! Just a friendly ping on this PR. Wanted to note that Issue #2554 was previously addressed by PR #2558 (closed by its author due to CI environment issues, not because of the fix itself). This PR picks up where that left off with the same approach — a minimal None guard in trimAndLoadJson that follows the existing error-handling pattern. The upstream CI failures (missing CONFIDENT_API_KEY, Black formatting) are pre-existing and unrelated to this change. Happy to address any review feedback!