AWS has launched Amazon Bedrock Advanced Prompt Optimization, a tool that automates prompt tuning and model migration on Amazon Bedrock. It accepts prompt templates in JSONL format with example inputs, ground truth answers, and evaluation metrics, then runs a feedback loop to iteratively rewrite and score prompts. Users can compare optimized prompts across up to 5 models simultaneously. Evaluation can be done via a custom AWS Lambda function (for concrete metrics like F1 or accuracy), LLM-as-a-Judge with a custom rubric (defaulting to Claude Sonnet 4.6), or free-form steering criteria. The tool supports multimodal inputs including PNG, JPG, and PDF. Results include evaluation scores, cost estimates, and latency data. It is available now across multiple AWS regions and billed at standard Bedrock inference token rates.
Sort: