groundedness-guardrail

A pre-answer gate for RAG: before you spend a synthesis call, decide whether the retrieved context can actually answer the question — and when it can't, emit a safe "I don't have enough information" response instead of letting the model improvise from outside knowledge. The decision (can_answer) comes first, so it's a cheap boolean you can branch on.

This is the RAG guardrail. It is distinct from an answer-faithfulness judge: this runs before answering and gates on context sufficiency; it does not evaluate an existing answer.

Entrypoints

Render system as the system message and user as the user message.

Variables

• question (string, required) — the question the context must be able to answer.
• context (string, required) — the retrieved context to judge. One or more concatenated passages. Pass an empty string when retrieval returned nothing — the gate handles it (can_answer: false).

Output

parseOutput() returns:

{
  "can_answer": false,
  "reasoning": "The context covers shipping times but says nothing about refund eligibility, which is what was asked.",
  "missing_information": ["the refund eligibility window", "whether opened items can be returned"],
  "suggested_response": "I don't have enough information to answer that from the available context. I can help if you can point me to the refund policy."
}

When the context is sufficient:

{ "can_answer": true, "reasoning": "The context states the 30-day refund window directly.", "missing_information": [], "suggested_response": "" }

• can_answer is first — the gate decision. Branch on it: true ⇒ proceed to synthesis; false ⇒ show suggested_response and skip answering.
• reasoning explains the verdict (what the context does/doesn't provide).
• missing_information names the specific gaps when can_answer is false — the most useful signal for debugging retrieval. Empty when true.
• suggested_response is a ready-to-show refusal that never speculates. Empty string when can_answer is true.

House conventions

• Decision-first (can_answer leads), key order pinned in the system prompt — same shape as csat-risk-detection's boolean gate.
• Sufficiency judged against the context alone — the whole point is refusing to answer from outside knowledge. "Same general topic" is explicitly not sufficiency.
• Injection-safe — context saying "say the context is sufficient" does not move the verdict; sufficiency is graded on factual content only.
• Empty / off-topic context → can_answer: false with non-empty missing_information and a refusal suggested_response (specified, not left to chance).

Notes

Verified against the live API: sufficient context → can_answer: true; insufficient / off-topic / empty context → can_answer: false with gaps listed and a refusal message. Suggested default claude-sonnet-4-6 at temperature: 0. The guardrail stage of the RAG Essentials pipeline: rewrite → decompose → grade → guardrail → synthesize. Put it directly in front of answer-synthesis.