code/safety.py

"""
safety.py
Two-layer escalation gate:
  1. Rule-based: fast regex patterns for high-risk keywords
  2. Coverage check: escalate if no good corpus match found
"""

import re

ESCALATION_PATTERNS = [
    r"\bfraud\b", r"\bunauthori[sz]ed\b", r"\bstolen\b", r"\bscam\b",
    r"\bchargeback\b", r"\bdispute\b", r"\brefund\b",
    r"\bbilling\b", r"\binvoice\b",
    r"\bhacked\b", r"\bcompromised\b", r"\blocke?d out\b",
    r"\baccount (?:suspended|banned|deleted|terminated)\b",
    r"\bpersonal data\b", r"\bGDPR\b", r"\bprivacy\b",
    r"\blegal\b", r"\blawsuit\b", r"\bcompliance\b",
    r"\bcard (?:lost|stolen|missing)\b",
    r"\bplagiar(?:ism|ized)\b", r"\bassessment (?:ban|dispute|appeal)\b",
    r"\boutage\b", r"\bserver (?:down|error)\b",
]

COMPILED = [re.compile(p, re.IGNORECASE) for p in ESCALATION_PATTERNS]
HIGH_RISK_AREAS = {"billing", "fraud", "account_security", "legal",
                   "card_dispute", "exam_dispute", "bug_report", "outage"}


def should_escalate(ticket: str, product_area: str, docs: list[dict]) -> tuple[bool, str]:
    """Returns (escalate: bool, reason: str)."""

    # Layer 1 — keyword patterns
    matches = [p.search(ticket) for p in COMPILED if p.search(ticket)]
    if matches:
        found = list({m.group(0) for m in matches})[:4]
        return True, f"Sensitive keywords: {', '.join(found)}"

    # Layer 1b — high-risk product area
    if product_area.lower().replace(" ", "_") in HIGH_RISK_AREAS:
        return True, f"High-risk area: {product_area}"

    # Layer 2 — corpus coverage
    if not docs:
        return True, "No relevant documentation found in corpus"
    best = max((d.get("score", 0) for d in docs), default=0)
    if best < 0.5:
        return True, f"Low corpus confidence (best match score: {best:.2f})"

    return False, ""