A pipeline relying on legacy text extraction doesn't just block machine learning; it inflates compute time by 10.9x and memory by 3.4x.