Amazon Bedrock Advanced Prompt Optimization now supports prompt shortening

Amazon Bedrock Advanced Prompt Optimization has introduced an optional prompt shortening feature that trims prompt length without degrading performance. Previously, optimization focused solely on improving evaluation scores, often resulting in longer prompts. The new setting addresses rising input token costs and latency concerns in large-scale or real-time applications by shortening prompts when beneficial.

When enabled, the optimizer evaluates whether a shorter version of a prompt maintains or improves evaluation scores compared to the original. If the shortened prompt meets these criteria, it replaces the original; otherwise, the original prompt is retained unchanged. This ensures no degradation in quality or increase in length, providing predictable outcomes for users.

The feature provides a compression ratio and token counts for both the original and optimized prompts in the output file and console. For instance, a compression ratio of 0.7 indicates a 30% reduction in tokens. This transparency helps users assess cost savings and performance trade-offs directly from the results.

Prompt shortening is compatible with all evaluation methods supported by Advanced Prompt Optimization, including steering criteria, LLM-as-a-judge rubrics, and custom Lambda evaluators. It is available in all AWS Regions where Advanced Prompt Optimization operates, and users can enable it through the prompt template settings in their input dataset.