SimulatedUserPersona
Simulated-user evaluation: an LLM role-plays a customer persona against a
real runtime, and an LLM judge scores the resulting transcript against a
rubric. This is the pre-deploy gate that scripted turn assertions
(EvalRunner) cannot provide — scripted turns test the happy path you
thought of; simulated users find the conversations you didn’t.
Properties
Section titled “Properties”goal: string;What they are trying to accomplish — the goal the judge scores against.
openingMessage?
Section titled “openingMessage?”optional openingMessage?: string;First user message. Generated from the persona when omitted.
profile
Section titled “profile”profile: string;Who the user is, e.g. “a busy parent in Colombo ordering a birthday cake”.
temperament?
Section titled “temperament?”optional temperament?: string;Behavioral style, e.g. “impatient; sends short fragmented messages; switches topics”.