Suppose your generative model needs to process very long inputs that exceed typical context limits. How would you approach managing or compressing that information effectively?