beta.chat.completions.parse returns unhandled ValidationError

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

In some occasions while using the Completion API with Structured Outputs, the SDK fails and returns a ValidationError:

ValidationError: 1 validation error for RawResponse
  Invalid JSON: EOF while parsing a value at line 1 column 600 [type=json_invalid, input_value='                        ...                       ', input_type=str]
    For further information visit https://errors.pydantic.dev/2.9/v/json_invalid

This does not happen every time, but we use it in a production service and this unpredictable behavior is hard to prevent.

To Reproduce

1. Create a Pydantic model
2. Instantiate an OpenAI client
3. Use the method OpenAI.beta.chat.completions.parse(...) with the following arguments
4. Repeat a few times for seeing the error

from pydantic import BaseModel
from openai import OpenAI

class RawResponse(BaseModel):
    answer: str

client = OpenAI(api_key=...)
completion = client.beta.chat.completions.parse(
                        model='gpt-4o-2024-08-06',
                        messages=messages,
                        max_tokens=750,
                        n=1,
                        stop=None,
                        temperature=0.1,
                        response_format=RawResponse
                    )

After a few times, this fails with:

ValidationError: 1 validation error for RawResponse
  Invalid JSON: EOF while parsing a value at line 1 column 600 [type=json_invalid, input_value='                        ...                       ', input_type=str]
    For further information visit https://errors.pydantic.dev/2.9/v/json_invalid

Code snippets

No response

OS

debian:bullseye-slim

Python version

CPython 3.10.8

Library version

openai 1.48.0

Start at OpenAI.beta.chat.completions.parse(...) and reproduce the intermittent Structured Outputs failure with the supplied Pydantic model. Trace where the response becomes RawResponse and inspect the error-handling boundary; done means the reproduction is covered by a regression check and no longer surfaces an unhandled ValidationError.