We haven't tried that, we might do that in the future. My intuition - not based on any research - is that recall should be a lot better from in context data vs

We haven't tried that, we might do that in the future. My intuition - not based on any research - is that recall should be a lot better from in context data vs. weights in the model. For our use case, precise recall is paramount.