IDP Accelerator - Auto Optimizer (Beta)

Tuning an intelligent document processing pipeline is expert, iterative work. Getting extraction accuracy from demo-good to production-good means rewriting prompts, reshaping output schemas, picking the right model per class, and adjusting classification strategy and OCR backends, then re-running evaluations to see whether each change actually helped. Auto Optimizer does that work autonomously. You supply a test set with ground truth, a maximum cost per page, and any optimization guidance you want respected. The agent handles the rest.

Under the hood, Auto Optimizer is an AI agent hosted on Amazon Bedrock AgentCore that drives your IDP Accelerator stack through purpose-built tools. It establishes a baseline evaluation, then loops: diagnose the weakest fields and documents, apply a targeted configuration change, re-evaluate against ground truth, and log its reasoning. Its decisions draw on 30 built-in expert skills covering prompt engineering, boolean and sparse-field extraction, nested schema design, multi-class setup, packet splitting, multilingual documents, token-limit and JSON-output failures, model selection, prompt caching, and conditional OCR cost reduction. Two guardrails are non-negotiable: the recommended configuration must come in at or under your cost-per-page ceiling, and the agent cannot touch your evaluation metrics. Accuracy is improved by extracting better, never by changing how it is scored.

You watch it work in the IDP Web UI. A live stream shows every agent turn and tool call as it happens, the agent's own optimization log explains what it tried and why, and the final report presents the recommended configuration alongside a table of every configuration tested, with accuracy, cost per page, and a one-click download of the archived YAML. Runs can be cancelled mid-flight without losing work and resumed later. Everything (the agent runtime, the test documents, the evaluations, the configurations) stays inside your own AWS account and region, so no document content or ground truth is shared with a third party. The result is a config you can review, diff, and deploy with confidence, produced in hours of unattended compute instead of weeks of manual tuning.