BadRequestError: Error code: 400

[powerhorse1986]

Have you searched existing issues? 🔎

• I have searched and found no existing issues

Desribe the bug

Dear Maarten,

How are you? Hope this message finds you all well.

Recently, I encountered a BadRequestError with Error code 400 from OpenAI when I run BERTopic with the representation_model being set to ChatGPT5.2. And the full error information was pasted as the following:

BadRequestError: Error code: 400 - {'error': {'message': "Unsupported parameter: 'stop' is not supported with this model.", 'type': 'invalid_request_error', 'param': 'stop', 'code': 'unsupported_parameter'}}

I did some research myself, and I figured out that the newer ChatGPT models, from some newer ChatGPT 4.x, no longer support "stop".

I solved this problem by copying and pasting your source code into my local Jupyter notebook, then I commented out the following two lines from bertopic.representation._openai:

if self.generator_kwargs.get("model"):
            self.model = generator_kwargs.get("model")
            del self.generator_kwargs["model"]
        if self.generator_kwargs.get("prompt"):
            del self.generator_kwargs["prompt"]
        #if not self.generator_kwargs.get("stop"):
        #    self.generator_kwargs["stop"] = "\n"

This fixed my problem immediately, however, I don't think this is the best solution. Because, some other users may still stick to those old models which still support "stop". I proposed the following code, you may want to consider to add into your source code:

First, we define a new method

def _safe_chat_completion(self, kwargs):
    """
    Try a chat completion call.
    If the model rejects the `stop` parameter, retry without it.
    """
    try:
        return client.chat.completions.create(**kwargs)

    except openai.BadRequestError as e:
        error_msg = str(e)

        # Explicitly detect the stop-parameter error
        if "Unsupported parameter: 'stop'" in error_msg:
            # Remove stop and retry once
            kwargs = dict(kwargs)      # shallow copy
            kwargs.pop("stop", None)

            return client.chat.completions.create(**kwargs)

        # Any other error should still raise
        raise

Remove the following block in extract_topics

if self.exponential_backoff:
    response = chat_completions_with_backoff(self.client, **kwargs)
else:
    response = self.client.chat.completions.create(**kwargs)

And call the new defined function:

if self.exponential_backoff:
    response = chat_completions_with_backoff(
        self.client,
        **kwargs
    )
else:
    response = self._safe_chat_completion(kwargs)

I tested it, it worked. Hope this might be helpful to you.

Reproduction

if self.generator_kwargs.get("model"):
            self.model = generator_kwargs.get("model")
            del self.generator_kwargs["model"]
        if self.generator_kwargs.get("prompt"):
            del self.generator_kwargs["prompt"]
        #if not self.generator_kwargs.get("stop"):
        #    self.generator_kwargs["stop"] = "\n"

BERTopic Version

0.17.4

[MaartenGr]

Thank you for the extensive description! I actually think it is alright to remove that line of code, especially since OpenAI is really focusing on people using their latest model and is a bit more aggressive in deprecating their older models this time around. If you want, a PR would be highly appreciated!

Small note that I'm trying to refactor the entire backend of BERTopic (see #2467), which is taking quite a bit of time. AI coding helps here but I want to check every single line of code for correctness. In that PR, I think I already removed that stop token and I might want to keep it that way.