improve anonymization of the names

processors/anonymizer.py

@@ -110,6 +110,41 @@ _PLACEHOLDERS = {
 }
 
 
+# Personal names that NER-based masking (Presidio/spaCy) commonly misses in customer-support chat:
+# agent sign-offs ("^Lynn", "*MelindaR", "- Laura :)", "-Dinesh") and greetings ("Hi Guy").
+# Measured on 119 annotated person names, Presidio alone left ~32 % of them visible, mostly exactly
+# these forms. The rules are deliberately narrow (a marker or a greeting is required) so that
+# ordinary capitalised words are not masked.
+_SIGNOFF_MARKER_RE = re.compile(r"(?<![\w])[\^~*]\s?[A-Z][A-Za-z]{1,14}(?![\w])")
+_SIGNOFF_DASH_RE = re.compile(r"(?<=\s)[-\u2013]\s?[A-Z][a-z]{1,14}(?=\s*(?::\)|\d+/\d+|[.!]?\s*$))")
+_GREETING_RE = re.compile(r"\b(Hi|Hello|Hey|Dear|Thanks|Thank you)(,?\s+)([A-Z][a-z]{1,14})\b(?!\s+[A-Z])")
+_NOT_A_NAME = {
+    "There", "All", "Team", "Everyone", "Everybody", "Again", "So", "Much", "You", "Guys", "Support", "Care", "Help",
+    "Sir", "Madam", "Customer", "Friend", "Folks", "Amazon", "Apple", "Uber", "Spotify",
+    "Do", "My", "Is", "It", "We", "The", "How", "What", "Please", "Sorry", "Yes", "No", "Ok", "Okay",
+}
+
+
+def mask_name_patterns(text: str) -> str:
+    """Mask agent sign-offs and greeted first names that NER-based masking tends to miss."""
+
+    def token_of(m: re.Match) -> str:
+        return re.sub(r"^[\^~*\-–]\s?", "", m.group(0))
+
+    def signoff(m: re.Match) -> str:
+        return m.group(0) if token_of(m) in _NOT_A_NAME else "[NAME]"
+
+    def dash(m: re.Match) -> str:
+        return m.group(0) if token_of(m) in _NOT_A_NAME else "- [NAME]"
+
+    def greeting(m: re.Match) -> str:
+        return m.group(0) if m.group(3) in _NOT_A_NAME else f"{m.group(1)}{m.group(2)}[NAME]"
+
+    text = _SIGNOFF_MARKER_RE.sub(signoff, text)
+    text = _SIGNOFF_DASH_RE.sub(dash, text)
+    return _GREETING_RE.sub(greeting, text)
+
+
 def _regex_anonymize(text: str) -> str:
     """Pure-regex anonymization for the formatted identifiers we can match safely."""
     text = _EMAIL_RE.sub("[EMAIL]", text)
@@ -142,7 +177,7 @@ def anonymize_text(text: str, language: str = "en") -> str:
                 "Refusing to process the text (set ANONYMIZER_ALLOW_FALLBACK=true only "
                 "for local development)."
             )
-        return _regex_anonymize(text)
+        return mask_name_patterns(_regex_anonymize(text))
 
     try:
         results = _analyzer.analyze(text=text, language=language)
@@ -158,14 +193,14 @@ def anonymize_text(text: str, language: str = "en") -> str:
             analyzer_results=results,
             operators=operators,
         )
-        return anonymized.text
+        return mask_name_patterns(anonymized.text)
     except Exception as e:  # pragma: no cover
         print(f"[anonymizer] Presidio failed at runtime: {e}")
         if not fallback_allowed():
             raise AnonymizationUnavailable(
                 f"Presidio failed on this text ({e}); refusing to fall back to regex-only masking."
             ) from e
-        return _regex_anonymize(text)
+        return mask_name_patterns(_regex_anonymize(text))
 
 
 def detect_pii(text: str, language: str = "en") -> list[dict[str, Any]]: