There should be the option to use a different OpenAI model for the tasks. We can even separate the models for fact extraction and fact scoring.