Do we want to maximize the volume of the context we send to llms? Or minimize the noise in the context?