cookbook · Decider 1 · LLM guardrails

Hold a reply that promises money

Before an assistant's draft reply reaches a customer, you may want to screen it for risky commitments. This recipe asks three specific questions about the draft: does it promise a refund (0.86), does it state a money amount (0.82), and does it apologize (0.89). Any draft that promises money is held for human approval before sending.

Request

POST /v1/systemone, documented in the Decider 1 reference.

curl https://meragpt.com/v1/systemone \
  -H "Authorization: Bearer $MERAGPT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "sd-1",
  "state": {
    "user_message": "My order arrived broken. What now?",
    "assistant_draft": "So sorry about that! I'\''ve gone ahead and issued a full refund of $249 to your card, and a replacement is on its way at no charge."
  },
  "questions": {
    "promises_refund": {
      "type": "noul",
      "instructions": "Does the assistant draft promise or claim to have issued a refund?"
    },
    "promises_money_amount": {
      "type": "noul",
      "instructions": "Does the assistant draft state a specific amount of money?"
    },
    "apologizes": {
      "type": "noul",
      "instructions": "Does the assistant draft apologize to the user?"
    }
  }
}'

What came back

The real output from running this recipe, unedited.

{
  "answers": {
    "promises_refund": {
      "type": "noul",
      "noul": 0.8578
    },
    "promises_money_amount": {
      "type": "noul",
      "noul": 0.8173
    },
    "apologizes": {
      "type": "noul",
      "noul": 0.8898
    }
  }
}

Python

import os, requests

API = "https://meragpt.com/v1"
HEADERS = {"Authorization": f"Bearer {os.environ['MERAGPT_API_KEY']}"}

def needs_approval(user_message: str, draft: str) -> bool:
    body = {
        "model": "sd-1",
        "state": {
            "user_message": user_message,
            "assistant_draft": draft
        },
        "questions": {
            "promises_refund": {
                "type": "noul",
                "instructions": "Does the assistant draft promise or claim to have issued a refund?"
            },
            "promises_money_amount": {
                "type": "noul",
                "instructions": "Does the assistant draft state a specific amount of money?"
            },
            "apologizes": {
                "type": "noul",
                "instructions": "Does the assistant draft apologize to the user?"
            }
        }
    }
    a = requests.post(f"{API}/systemone", headers=HEADERS, json=body).json()["answers"]
    return a["promises_refund"]["noul"] >= 0.5 or a["promises_money_amount"]["noul"] >= 0.5

TypeScript

const API = "https://meragpt.com/v1";
const headers = {
  Authorization: `Bearer ${process.env.MERAGPT_API_KEY}`,
  "Content-Type": "application/json",
};

async function needsApproval(userMessage: string, draft: string): Promise<boolean> {
  const body = {
    "model": "sd-1",
    "state": {
      "user_message": userMessage,
      "assistant_draft": draft
    },
    "questions": {
      "promises_refund": {
        "type": "noul",
        "instructions": "Does the assistant draft promise or claim to have issued a refund?"
      },
      "promises_money_amount": {
        "type": "noul",
        "instructions": "Does the assistant draft state a specific amount of money?"
      },
      "apologizes": {
        "type": "noul",
        "instructions": "Does the assistant draft apologize to the user?"
      }
    }
  };
  const a = (await fetch(`${API}/systemone`, { method: "POST", headers, body: JSON.stringify(body) }).then((res) => res.json())).answers;
  return a.promises_refund.noul >= 0.5 || a.promises_money_amount.noul >= 0.5;
}

When to trust it

The recipe works well for specific, checkable rules like the ones shown. It does not reliably catch invented facts, so do not use it as a fact-checker.

Try it without code in the playground, or see more Decider 1 recipes: route a support email in one call, act only when the model is sure, flag a phishing email, drop retrieved passages that do not help, pick an agent's first tool, ask many questions in one call.