I’m a bit befuddled why this is not enabled by default. Customers of managed agent services don’t want Assistants to use generated text 100% of the time, we occasionally want the Assistant to respond with precise, pre-provided language, particularly after tool calls.
For example, if a customer interacts with my Assistant to purchase a product, I want the Assistant to respond with exact and consistent language after calling the tool to make the purchase. “Your purchase has been processed and will be available 5 day etc…”. Or, maybe A Canadian Airline wants to be able to insert specific language on their airfare refund policy when the user asks about refunds, instead of letting the model generate it. Including this language in the system message or other context doesn’t work because the Assistant may mis-state it or forget to respond with it.
Someone said on another thread this was a tradeoff for Assistants. I respectfully disagree - being able to control when your Assistant uses exact language in certain scenarios, vs. generated responses, seems like an important feature in any managed agent service.