Question

A Generative AI Engineer is building a customer-facing HR benefits assistant using a Databricks Foundation Model API chat endpoint. The business requires that the assistant only answer questions about company benefits, always respond in a professional and empathetic tone, and never speculate about individual medical conditions. The engineer wants to enforce these behavioral constraints in the most maintainable way, without hardcoding them into every user turn. Which prompt engineering approach best satisfies these requirements?

Accepted Answer

Define the constraints once in a system-role message that persists across the conversation, defining the assistant's persona, scope, and prohibitions.

Answer

Prepend the constraints to each user message string before sending it as a single user-role message.

Answer

Fine-tune the foundation model on transcripts so the constraints are learned into the model weights and never need to be specified at inference.

Answer

Add the constraints as a few-shot set of example user/assistant pairs appended after every user question at runtime.