Decoupling Reasoning from Observations for Efficient Augmented Language Models - View it on GitHub