Aug 5 - Update code to follow up ST - MT14c

app/routes/support.py

@@ -20,6 +20,38 @@ from app.utils.notifications import notify
 bp = Blueprint('support', __name__, url_prefix='/support')
 logger = logging.getLogger(__name__)
 
+
+# ── Outbound PII redaction ────────────────────────────────────────────────────
+# Groq is a THIRD PARTY. Customers routinely paste contact details (their own,
+# or a coworker's) into a support question, and none of that needs to leave the
+# app to get a helpful, generic answer. This scrubs a best-effort set of PII
+# patterns from the copy of the text sent to Groq ONLY — the original is still
+# stored verbatim in support_chat_messages, so the customer's own conversation
+# history reads normally in the app and staff see what was actually said.
+#
+# Best-effort by design: over-redacting a support question costs nothing, while
+# under-redacting leaks a real address. Order matters — the 13–19 digit card
+# pattern runs before the phone pattern so a card number is not partly consumed
+# as a phone number first.
+import re as _re
+
+_PII_PATTERNS = [
+    (_re.compile(r'[\w.+-]+@[\w-]+\.[\w.-]+'), '[redacted-email]'),
+    (_re.compile(r'\b\d{3}-\d{2}-\d{4}\b'), '[redacted-ssn]'),
+    (_re.compile(r'\b(?:\d[ -]?){13,19}\b'), '[redacted-number]'),
+    (_re.compile(r'\b(?:\+?1[ .-]?)?\(?\d{3}\)?[ .-]?\d{3}[ .-]?\d{4}\b'), '[redacted-phone]'),
+]
+
+
+def _redact_pii(text):
+    """Best-effort scrub of email/phone/SSN/card-like sequences from outbound text."""
+    if not text:
+        return text
+    redacted = text
+    for pattern, placeholder in _PII_PATTERNS:
+        redacted = pattern.sub(placeholder, redacted)
+    return redacted
+
 # ── Groq system prompt ────────────────────────────────────────────────────────
 
 _SYSTEM_PROMPT = """\
@@ -159,9 +191,11 @@ def chat_message():
         client = Groq(api_key=api_key)
 
         messages = [{'role': 'system', 'content': _system_prompt_with_kb()}]
+        # Redact before the text leaves the app for Groq. The unredacted
+        # originals are persisted below, so nothing is lost in-app.
         for m in prior[-20:]:
-            messages.append({'role': m.role, 'content': m.content})
-        messages.append({'role': 'user', 'content': user_message})
+            messages.append({'role': m.role, 'content': _redact_pii(m.content)})
+        messages.append({'role': 'user', 'content': _redact_pii(user_message)})
 
         model = os.environ.get('GROQ_MODEL', 'llama-3.3-70b-versatile')
         completion = client.chat.completions.create(