How do you stop RAG from making up missing steps?

I built an internal assistant for HR and ops docs. The biggest headaches happen when an employee asks about a policy that isn't fully written down yet. Instead of just stopping, the model tries to be "helpful" and fabricates logical-sounding (but completely wrong) company steps.

I’ve tweaked system prompts a hundred times ("if unsure, say I don't know"), but it still slips up. How are you strictly locking down responses when the retrieved context is incomplete?