Problem with system prompt when using Mode.JSON

[dimentary]

Is your feature request related to a problem? Please describe.
I am using LLM for roleplay storywriting and also use mode.JSON for the models that don't support tool calling. When using this mode, InstructorAI inserts this text into the system prompt:

As a genius expert, your task is to understand the content and provide
the parsed objects in json that match the following json_schema:

{json.dumps(response_model.model_json_schema(), indent=2, ensure_ascii=False)}

Make sure to return an instance of the JSON, not the schema itself

The problem is the "As a genius expert" part. I understand that it might improve the quality of the LLM outputs, but in my case it breaks character consistency as now it acts as a "genius expert".

Describe the solution you'd like
A more generic system prompt for the JSON mode.

Describe alternatives you've considered
An ability to pass custom system prompt by the user.

[sftblw]

I had a similar problem, and monkey-patched my system prompt via the hook system.

def _instructor_hook_system_message_override(*args, **kwargs):

    if '___override_system_message' in kwargs:
        kwargs["messages"][0]["content"] = kwargs['___override_system_message']

    if '___override_system_message' in kwargs:
        del kwargs['___override_system_message']

instructor_client.on("completion:kwargs", _instructor_hook_system_message_override)

# then

kwargs = {
    "___override_system_message": "Source system message without any modifications from the instructor framework",
}
instructor_client.chat.completions.create_with_completion(
    #...other_args...,
    **kwargs
)

instructor/instructor/process_response.py

Lines 444 to 459 in df62492

	if new_kwargs["messages"][0]["role"] != "system":
	new_kwargs["messages"].insert(
	0,
	{
	"role": "system",
	"content": message,
	},
	)
	elif isinstance(new_kwargs["messages"][0]["content"], str):
	new_kwargs["messages"][0]["content"] += f"\n\n{message}"
	elif isinstance(new_kwargs["messages"][0]["content"], list):
	new_kwargs["messages"][0]["content"][0]["text"] += f"\n\n{message}"
	else:
	raise ValueError(
	"Invalid message format, must be a string or a list of messages"
	)

(It seems to have been refactored (moved) on the main branch into this file:

instructor/instructor/providers/openai/utils.py

Lines 436 to 451 in 5a7ba3d

	if new_kwargs["messages"][0]["role"] != "system":
	new_kwargs["messages"].insert(
	0,
	{
	"role": "system",
	"content": message,
	},
	)
	elif isinstance(new_kwargs["messages"][0]["content"], str):
	new_kwargs["messages"][0]["content"] += f"\n\n{message}"
	elif isinstance(new_kwargs["messages"][0]["content"], list):
	new_kwargs["messages"][0]["content"][0]["text"] += f"\n\n{message}"
	else:
	raise ValueError(
	"Invalid message format, must be a string or a list of messages"
	)

)

In JSON and JSON_SCHEMA modes, it seems to be already sending the schema via response_format. Someone like me would want to modify or remove the "genius" prompt. It would be appreciated if there were a way to customize that prompt.

[Pringled]

+1, it would be great to have the ability to pass a custom prompt here; the "genius expert" can have a big effect on the output style.