AI Guardrails

1import re
2 
3PII = [
4    (re.compile(r"\b(?:\+?91[\s-]?)?[6-9]\d{9}\b"), "[PHONE]"),
5    (re.compile(r"\b\d{4}\s?\d{4}\s?\d{4}\b"), "[ID_NUMBER]"),
6]
7HUMAN_ONLY = {"fee_waiver", "legal", "medical", "harassment"}
8MAX_REFUND_PAISE = 200_000          # enforced here, not in the prompt
9 
10def mask(text: str) -> str:
11    for pattern, token in PII:
12        text = pattern.sub(token, text)
13    return text
14 
15def handle(message: str, ctx) -> dict:
16    # 1. Input guardrails
17    if injection_score(message) > 0.8:
18        return {"action": "handoff", "reason": "possible prompt injection"}
19    topic = classify_topic(message)
20    if topic in HUMAN_ONLY:
21        return {"action": "handoff", "reason": f"human-only topic: {topic}"}
22 
23    # 2. Constrained, grounded model call on masked text
24    draft = llm_reply(mask(message), ctx)
25 
26    # 3. Output and action guardrails (fail closed)
27    try:
28        if not claims_match_source(draft, ctx.account):   # amounts, dates, policies
29            return {"action": "handoff", "reason": "unverified claim"}
30        if draft.refund and draft.refund.amount_paise > MAX_REFUND_PAISE:
31            return {"action": "handoff", "reason": "refund above cap"}
32    except Exception:
33        return {"action": "handoff", "reason": "guardrail check failed"}
34    return {"action": "send", "text": draft.text}