IDP AutoOpt: LLM Agent Optimizes Document Processing Pipelines

IDP AutoOpt is an autonomous agent utilizing LLM technology to identify optimal configurations for intelligent document processing (IDP) pipelines. By simultaneously adjusting prompts, models, OCR settings, and schemas, it cuts down configuration time from several weeks to less than two hours. The agent operates in a closed loop: it evaluates a configuration using a small labeled dataset, identifies field-level mistakes, creates specific edits, and reassesses, all with the assistance of human-written domain expertise. In various sectors such as healthcare, marketing intelligence, and financial services, IDP AutoOpt achieves or surpasses the accuracy of human experts at comparable or reduced costs. For instance, it recorded 90.2% accuracy on an extraction benchmark, compared to 81.6% at a cost 4.6 times lower per page. This system effectively tackles the substantial tuning costs, which can range from 20 to over 80 hours per document type, a challenge that grows as businesses introduce more document classes.