Characterizing Prompt Compression Methods for Long Context Inference