Yep! We're also experimenting with small models to filter out sections that are irrelevant for the agent given its current goal/task