Question

A team runs a production customer-support chatbot on Amazon Bedrock using a specific foundation model version. Bedrock announces a newer model version, and the team wants to adopt it. Leadership requires that the new model be validated against real production traffic patterns without exposing customers to any degraded responses before a full cutover. Which approach best satisfies this requirement while allowing objective comparison of output quality?

Accepted Answer

Run the new model in shadow mode by asynchronously sending a mirrored copy of production requests to it, log both models' responses, and compare quality offline before any customer-facing switch.

Answer

Deploy the new model version directly to 100% of traffic and monitor CloudWatch invocation error metrics for a rollback trigger.

Answer

Create a Bedrock automatic model evaluation job using the built-in datasets to score the new model, then cut over once scores are acceptable.

Answer

Enable a Bedrock Guardrail contextual grounding filter on the new model and cut over immediately, relying on the filter to block any bad responses.