[bug] [langchain] JSON serialization fails with non-string dictionary keys (e.g., Azure PISubType enum)

[seanlee10]

Describe the bug
The LangChain tracer fails to serialize span attributes when dictionary keys are non-string types (such as enum values). The safe_json_dumps fallback only handles non-serializable values via default=str, but does not handle non-string dictionary keys, which causes json.dumps() to fail.

To Reproduce

1. Use Azure AI Document Intelligence to process a document with PII detection enabled
2. Pass the resulting document through a LangChain chain with openinference-instrumentation-langchain enabled
3. The tracer fails when attempting to serialize metadata containing PISubType enum keys

Expected behavior
The tracer should successfully serialize objects with non-string dictionary keys by converting them to strings, rather than raising a TypeError.

Screenshots

2025-12-08 14:48:07.927 | ERROR | Failed to get attribute.
Traceback (most recent call last):
  File ".../openinference/instrumentation/langchain/_tracer.py", line 481, in _json_dumps
    return json.dumps(obj, cls=_OpenInferenceJSONEncoder, ensure_ascii=False)
  ...
TypeError: keys must be str, int, float, bool or None, not PISubType

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File ".../openinference/instrumentation/langchain/_tracer.py", line 327, in wrapper
    yield from wrapped(*args, **kwargs)
  File ".../openinference/instrumentation/langchain/_tracer.py", line 421, in _convert_io
    json_value = _json_dumps(obj)
  File ".../openinference/instrumentation/langchain/_tracer.py", line 484, in _json_dumps
    return safe_json_dumps(obj)
  File ".../openinference/instrumentation/helpers.py", line 21, in safe_json_dumps
    return json.dumps(obj, default=str, ensure_ascii=False, **kwargs)
  ...
TypeError: keys must be str, int, float, bool or None, not PISubType

Desktop (please complete the following information):

• OS: Linux
• Python: 3.11

Additional context
Add any other context about the problem here.

[dosubot]

The serialization fails because Python's json.dumps requires dictionary keys to be strings (or int, float, bool, None), but the tracer passes through dicts with enum keys (like Azure PISubType), which triggers a TypeError. The current safe_json_dumps helper only stringifies values, not keys, so it doesn't prevent this error (source).

To fix this, you can add a recursive utility that walks any mapping and converts all dictionary keys to strings before serialization. For example:

def stringify_keys(obj):
    if isinstance(obj, dict):
        return {str(k): stringify_keys(v) for k, v in obj.items()}
    elif isinstance(obj, list):
        return [stringify_keys(i) for i in obj]
    else:
        return obj

You'd call this on the object before passing it to json.dumps in safe_json_dumps or just before serialization in the tracer (e.g., in _convert_io or _metadata). This ensures all keys are valid for JSON, even if they're enums or other non-string types (source).

Trade-off: this approach loses the original type of the keys (e.g., you can't distinguish between an enum and its string value after serialization), but it's necessary for JSON compatibility and is a common practice.

There are no recent PRs or built-in utilities in the codebase that address this yet, so a local patch or monkey-patch to safe_json_dumps is the most direct workaround for now.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}