Guardrails for avoiding harmful medical product recommendations and off-label promotion in generative AI models