How to ask full forecasting questions?

[dpaleka]

I want to run the forecaster on some project of mine. Concretely, I have questions with fields (title, resolution_criteria, resolution_date and created_date) (similar to what Metaculus API has); and I want to instantiate your forecaster to give a forecast on this question, if the forecaster was run e.g. on forecast_date (say 1 May 2024).

What is the data one needs to send to the forecasting_search_local endpoint? Could you perhaps provide an example?

[justinphan3110cais]

Hi.

async def forecasting_search_local(data: dict) -> str:
    model = data['model']
    messages = data['messages']
    breadth = data.get('breadth')
    plannerPrompt = data.get('plannerPrompt')
    factorized_prompt = data.get('factorizedPrompt')
    publisherPrompt = data.get('publisherPrompt')
    search_type = data.get('search_type')
    before_timestamp = data.get('beforeTimestamp')

    multi_agents = ForecastingMultiAgents(model, 
                                          breadth, 
                                          plannerPrompt, 
                                          publisherPrompt, 
                                          search_type, 
                                          before_timestamp,
                                          factorized_prompt)

    response = multi_agents.completions(messages)
    
    # Collect the full response
    full_response = ""
    async for chunk in response:
        full_response += chunk
    return full_response

For the function above you only need to worry about the messages which is something similiar to OpenAI's api input:
[dict(role="user", content=question)]. We use breadth from 5 to 7, you can find the prompts in prompts.py which are the default value if you not passing anything. beforeTimestamp is where you limit the news article for the sources