A thing worth thinking about before you put a compression step in the middle of a pipeline: lossy summarisation is silent, and everything downstream treats what it receives as complete. Compressing a hundred-page contract to two pages is genuinely useful for a human who will go back to the source. It is a different proposition when the two pages are the only thing a downstream model ever sees, because the model has no way to know that the indemnity cap was in a footnote that did not survive. It will answer the question confidently either way. What I would want is not better compression. It is for the compressor to emit what it dropped, or at minimum a confidence that the answer to a given question is still recoverable from the compressed form.