Write a RAG Prompt That Cites and Refuses
Explore how to design RAG prompts that ground AI responses strictly in retrieved policy texts by requiring exact quotations and managing unsupported claims with refusal rules. Understand output contracts that enable automatic validation and safe escalation to human review, ensuring trustworthy and auditable support workflows.
A support draft can look polished while still being wrong. When retrieval returns irrelevant internal policy text, a generic prompt often produces a confident reply that invents a refund exception such as We can approve refunds after 60 days if the item is unopened even when no retrieved chunk contains that rule.
With that failure in view, the engineering requirement becomes concrete. Generation must treat the retrieved chunks as the only allowable source of policy claims, and anything that cannot be pointed to in those chunks is unsupported and must not appear. The focus here is prompt template structure, not model choice.
To spot which instructions actually enforce grounding, we can compare a generic answer prompt to a RAG prompt skeleton that includes quoting, citations, and refusal rules.
A prompt contract that blocks invented policy
With the grounding ...