tool_calls and other keys in OpenAI API return value set to None are stripped

[gaborvar]

Hi @pamelafox
I use message_builder in code that relies on the tool_calls feature of GPT.

model_helper.py assumes that certain keys (if present) cannot be set to None in the return value of the ChatCompletion API:

openai-messages-token-helper/src/openai_messages_token_helper/model_helper.py

Line 124 in 5c579fd

raise ValueError(f"Could not encode unsupported message value type: {type(value)}")

Example return value:

ChatCompletionMessage(content='Inform, assist, connect with legal experts, facilitate paperwork.', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=None)

Keys set to None are routinely returned by the model, in my case gpt-4o 2024-08-06 .

It seems to be a bug which is suppressed in other parts of the code if tool_calls are not used. I.e. the code omits None keys before they hit this code line. However, for tool_calls to work correctly I need to pass back to the model what it has returned previously. OpenAI API returns an error if I omit certain parts of its previous returns.

Is tool_calls and openai_messages_token_helper together a niche scenario or is it worth fixing in the repo?

See also #21 #20
Thanks