Add system_prompt param to GeneralLLM.invoke function

[Stepka]

This PR adds a system_prompt parameter to the GeneralLLM.invoke function.

The main motivation is to remove the need to rely on the deprecated Perplexity class for setting system-level instructions. With this change, system prompts can be passed directly to GeneralLLM.invoke, making the API more flexible and future-proof.

This is particularly useful for existing implementations such as Q3TemplateBot2024, where system prompts are currently handled via the deprecated Perplexity abstraction.

Changes

• Added optional system_prompt parameter to GeneralLLM.invoke

• Allows explicit system-level instructions without using deprecated classes

• Improves compatibility with current and future LLM backends

Backward compatibility

• Existing calls to GeneralLLM.invoke remain unchanged

• The new parameter is optional and does not break current usage

Motivation

Deprecating Perplexity requires an alternative way to inject system prompts. This change provides a minimal and clean solution while keeping the existing API stable.

Example

This implementation of Q3TemplateBot2024.run_research

    async def run_research(self, question: MetaculusQuestion) -> str:
        system_prompt = clean_indents(
            """
            You are an assistant to a superforecaster.
            The superforecaster will give you a question they intend to forecast on.
            To be a great assistant, you generate a concise but detailed rundown of the most relevant news, including if the question would resolve Yes or No based on current information.
            You do not produce forecasts yourself.
            """
        )

        # Note: The original q3 bot did not set temperature, and I could not find the default temperature of perplexity
        response = await Perplexity(
            temperature=0.1, system_prompt=system_prompt
        ).invoke(question.question_text)
        return response

can be changed to

    async def run_research(self, question: MetaculusQuestion) -> str:
        system_prompt = clean_indents(
            """
            You are an assistant to a superforecaster.
            The superforecaster will give you a question they intend to forecast on.
            To be a great assistant, you generate a concise but detailed rundown of the most relevant news, including if the question would resolve Yes or No based on current information.
            You do not produce forecasts yourself.
            """
        )

        # Note: The original q3 bot did not set temperature, and I could not find the default temperature of perplexity
        model = self.get_llm("researcher", "llm") # Assume there is a `perplexity/sonar-pro` model
        response = await model.invoke(question.question_text, system_prompt=system_prompt)
        return response