Question

A retail company is deploying a customer-facing agent built in Copilot Studio that answers product warranty questions grounded on the company's knowledge base. During testing, the agent frequently invents warranty terms not present in the source documents and sometimes ignores the grounding data entirely. As the solution architect defining prompt-engineering guidelines for the build team, which approach should you prioritize to most directly reduce these fabricated answers?

Accepted Answer

Add explicit grounding instructions to the system prompt that require the agent to answer only from retrieved knowledge sources and to state when information is unavailable.

Answer

Increase the model's temperature setting so responses are more varied and natural-sounding to customers.

Answer

Switch the underlying model to a larger foundation model with a bigger parameter count.

Answer

Remove the knowledge base grounding and let the model rely on its pretrained knowledge for faster responses.