Extract hardcoded confidence thresholds to named constants

src/document_anonymizer/api/schemas.py

@@ -3,6 +3,7 @@
 from pydantic import BaseModel, Field
 
 from document_anonymizer.anonymization.strategies import AnonymizationStrategy
+from document_anonymizer.constants import DEFAULT_SCORE_THRESHOLD
 
 _EXAMPLE_TEXT = (
     "Herr Max Mustermann, geboren am 15.03.1985, "
@@ -26,10 +27,10 @@ class DetectionRequest(BaseModel):
         examples=["de"],
     )
     score_threshold: float = Field(
-        default=0.35,
+        default=DEFAULT_SCORE_THRESHOLD,
         ge=0.0,
         le=1.0,
-        examples=[0.35],
+        examples=[DEFAULT_SCORE_THRESHOLD],
     )
 
 
@@ -74,10 +75,10 @@ class AnonymizeRequest(BaseModel):
         examples=[{"PERSON": "fake", "DE_IBAN": "mask"}],
     )
     score_threshold: float = Field(
-        default=0.35,
+        default=DEFAULT_SCORE_THRESHOLD,
         ge=0.0,
         le=1.0,
-        examples=[0.35],
+        examples=[DEFAULT_SCORE_THRESHOLD],
     )
 
 

src/document_anonymizer/constants.py

@@ -0,0 +1,24 @@
+"""Shared constants for confidence scoring and detection thresholds."""
+
+__all__ = [
+    "DEFAULT_SCORE_THRESHOLD",
+    "RECOGNIZER_BASE_SCORE_HIGH",
+    "RECOGNIZER_BASE_SCORE_LOW",
+    "TIER_HIGH_THRESHOLD",
+    "TIER_MEDIUM_THRESHOLD",
+]
+
+# Minimum confidence score for a detection to be included in results.
+# Lower values catch more entities but increase false positives.
+DEFAULT_SCORE_THRESHOLD = 0.35
+
+# Tier boundaries for the entity review panel.
+# High-confidence entities (>= TIER_HIGH_THRESHOLD) are shown expanded by default.
+TIER_HIGH_THRESHOLD = 0.7
+TIER_MEDIUM_THRESHOLD = 0.5
+
+# Base confidence scores for pattern-based recognizers.
+# Higher base = more confident the pattern alone is a real match.
+# Context words boost these scores at runtime.
+RECOGNIZER_BASE_SCORE_HIGH = 0.5
+RECOGNIZER_BASE_SCORE_LOW = 0.3

src/document_anonymizer/detection/recognizers/german_date.py

@@ -8,6 +8,8 @@
 
 from presidio_analyzer import Pattern, PatternRecognizer
 
+from document_anonymizer.constants import RECOGNIZER_BASE_SCORE_LOW
+
 # DD.MM.YYYY — standard German date format
 _DATE_PATTERN = r"\b(?:0[1-9]|[12]\d|3[01])\.(?:0[1-9]|1[0-2])\.\d{4}\b"
 
@@ -37,7 +39,7 @@ class GermanDateRecognizer(PatternRecognizer):
 
     def __init__(self) -> None:
         patterns = [
-            Pattern("german_date_full", _DATE_PATTERN, 0.3),
+            Pattern("german_date_full", _DATE_PATTERN, RECOGNIZER_BASE_SCORE_LOW),
             Pattern("german_date_short", _DATE_SHORT_PATTERN, 0.2),
         ]
         super().__init__(

src/document_anonymizer/detection/recognizers/german_handelsreg.py

@@ -8,6 +8,8 @@
 
 from presidio_analyzer import Pattern, PatternRecognizer
 
+from document_anonymizer.constants import RECOGNIZER_BASE_SCORE_HIGH
+
 _HANDELSREG_PATTERN = r"\bHR[AB]\s?\d{3,6}\s?[A-Z]?\b"
 
 _CONTEXT_WORDS = [
@@ -30,7 +32,11 @@ class GermanHandelsregisterRecognizer(PatternRecognizer):
 
     def __init__(self) -> None:
         patterns = [
-            Pattern("german_handelsregister", _HANDELSREG_PATTERN, 0.5),
+            Pattern(
+                "german_handelsregister",
+                _HANDELSREG_PATTERN,
+                RECOGNIZER_BASE_SCORE_HIGH,
+            ),
         ]
         super().__init__(
             supported_entity="DE_HANDELSREGISTER",

src/document_anonymizer/detection/recognizers/german_iban.py

@@ -4,6 +4,8 @@
 
 from presidio_analyzer import Pattern, PatternRecognizer
 
+from document_anonymizer.constants import RECOGNIZER_BASE_SCORE_HIGH
+
 # DE + 2 check digits + 18 digits (bank code + account number)
 # Allows optional spaces every 4 characters
 _IBAN_PATTERN = r"\bDE\d{2}\s?\d{4}\s?\d{4}\s?\d{4}\s?\d{4}\s?\d{2}\b"
@@ -42,7 +44,7 @@ class GermanIbanRecognizer(PatternRecognizer):
     """Detects German IBANs with context boosting and checksum validation."""
 
     ENTITIES: ClassVar[list[str]] = ["DE_IBAN"]
-    DEFAULT_SCORE = 0.5
+    DEFAULT_SCORE = RECOGNIZER_BASE_SCORE_HIGH
 
     def __init__(self) -> None:
         patterns = [

src/document_anonymizer/detection/recognizers/german_id_card.py

@@ -13,6 +13,8 @@
 
 from presidio_analyzer import Pattern, PatternRecognizer
 
+from document_anonymizer.constants import RECOGNIZER_BASE_SCORE_LOW
+
 # Restricted alphanumeric: letters that cannot be confused with digits
 _VALID_LETTERS = "CFGHJKLMNPRTVWXYZ"
 
@@ -63,7 +65,7 @@ class GermanIdCardRecognizer(PatternRecognizer):
 
     def __init__(self) -> None:
         patterns = [
-            Pattern("german_id_card", _ID_PATTERN, 0.3),
+            Pattern("german_id_card", _ID_PATTERN, RECOGNIZER_BASE_SCORE_LOW),
         ]
         super().__init__(
             supported_entity="DE_ID_CARD",

src/document_anonymizer/detection/recognizers/german_phone.py

@@ -4,6 +4,11 @@
 
 from presidio_analyzer import Pattern, PatternRecognizer
 
+from document_anonymizer.constants import (
+    RECOGNIZER_BASE_SCORE_HIGH,
+    RECOGNIZER_BASE_SCORE_LOW,
+)
+
 # International format: +49 followed by area code and number
 _INTL_PATTERN = r"\+49\s?\(?\d{2,4}\)?\s?\d{3,8}(?:[\s-]?\d{1,5})?\b"
 
@@ -37,8 +42,10 @@ class GermanPhoneRecognizer(PatternRecognizer):
 
     def __init__(self) -> None:
         patterns = [
-            Pattern("german_phone_intl", _INTL_PATTERN, 0.5),
-            Pattern("german_phone_domestic", _DOMESTIC_PATTERN, 0.3),
+            Pattern("german_phone_intl", _INTL_PATTERN, RECOGNIZER_BASE_SCORE_HIGH),
+            Pattern(
+                "german_phone_domestic", _DOMESTIC_PATTERN, RECOGNIZER_BASE_SCORE_LOW
+            ),
             Pattern("german_phone_mobile", _MOBILE_PATTERN, 0.6),
         ]
         super().__init__(

src/document_anonymizer/detection/recognizers/german_tax.py

@@ -4,6 +4,8 @@
 
 from presidio_analyzer import Pattern, PatternRecognizer
 
+from document_anonymizer.constants import RECOGNIZER_BASE_SCORE_LOW
+
 # Steuer-ID: exactly 11 digits, no leading zero
 _STEUER_ID_PATTERN = r"\b[1-9]\d{10}\b"
 
@@ -67,12 +69,14 @@ def __init__(self) -> None:
             Pattern(
                 "steuer_id",
                 _STEUER_ID_PATTERN,
-                0.3,  # Low base — 11 digits are common; context boosts it
+                # Low base — 11 digits are common; context boosts it
+                RECOGNIZER_BASE_SCORE_LOW,
             ),
             Pattern(
                 "steuernummer",
                 _STEUERNUMMER_PATTERN,
-                0.3,  # Below default threshold; requires context words
+                # Below default threshold; requires context words
+                RECOGNIZER_BASE_SCORE_LOW,
             ),
         ]
         super().__init__(

src/document_anonymizer/document/pdf_handler.py

@@ -15,6 +15,7 @@
 
 from document_anonymizer.anonymization.engine import anonymize_text
 from document_anonymizer.anonymization.strategies import AnonymizationStrategy
+from document_anonymizer.constants import DEFAULT_SCORE_THRESHOLD
 from document_anonymizer.document.text_handler import detect_pii_in_text
 
 if TYPE_CHECKING:
@@ -82,7 +83,7 @@ def detect_pii_in_pdf(
     analyzer: AnalyzerEngine,
     pdf_bytes: bytes,
     language: str = "de",
-    score_threshold: float = 0.35,
+    score_threshold: float = DEFAULT_SCORE_THRESHOLD,
 ) -> list[PdfDetection]:
     """Detect PII in a PDF with page and position metadata.
 
@@ -136,7 +137,7 @@ def redact_pdf(
     analyzer: AnalyzerEngine,
     pdf_bytes: bytes,
     language: str = "de",
-    score_threshold: float = 0.35,
+    score_threshold: float = DEFAULT_SCORE_THRESHOLD,
 ) -> tuple[bytes, list[PdfDetection]]:
     """Physically redact PII from a PDF document.
 
@@ -284,7 +285,7 @@ def anonymize_pdf_text(
     pdf_bytes: bytes,
     strategy: AnonymizationStrategy = AnonymizationStrategy.REPLACE,
     language: str = "de",
-    score_threshold: float = 0.35,
+    score_threshold: float = DEFAULT_SCORE_THRESHOLD,
 ) -> tuple[str, list[RecognizerResult]]:
     """Extract text from PDF, detect and anonymize PII.
 

src/document_anonymizer/document/processor.py

@@ -4,6 +4,7 @@
 from presidio_anonymizer import AnonymizerEngine
 
 from document_anonymizer.anonymization.strategies import AnonymizationStrategy
+from document_anonymizer.constants import DEFAULT_SCORE_THRESHOLD
 from document_anonymizer.document.pdf_handler import PdfDetection, redact_pdf
 from document_anonymizer.document.text_handler import anonymize_plain_text
 
@@ -15,7 +16,7 @@ def process_text(
     strategy: AnonymizationStrategy = AnonymizationStrategy.REPLACE,
     entity_strategies: dict[str, AnonymizationStrategy] | None = None,
     language: str = "de",
-    score_threshold: float = 0.35,
+    score_threshold: float = DEFAULT_SCORE_THRESHOLD,
 ) -> tuple[str, list[RecognizerResult]]:
     """Process plain text: detect and anonymize PII."""
     return anonymize_plain_text(
@@ -33,7 +34,7 @@ def process_pdf(
     analyzer: AnalyzerEngine,
     pdf_bytes: bytes,
     language: str = "de",
-    score_threshold: float = 0.35,
+    score_threshold: float = DEFAULT_SCORE_THRESHOLD,
 ) -> tuple[bytes, list[PdfDetection]]:
     """Process PDF: detect PII and apply physical redaction."""
     return redact_pdf(

src/document_anonymizer/document/text_handler.py

@@ -5,6 +5,7 @@
 
 from document_anonymizer.anonymization.engine import anonymize_text
 from document_anonymizer.anonymization.strategies import AnonymizationStrategy
+from document_anonymizer.constants import DEFAULT_SCORE_THRESHOLD
 
 
 def _deduplicate_overlapping(
@@ -39,7 +40,7 @@ def detect_pii_in_text(
     engine: AnalyzerEngine,
     text: str,
     language: str = "de",
-    score_threshold: float = 0.35,
+    score_threshold: float = DEFAULT_SCORE_THRESHOLD,
 ) -> list[RecognizerResult]:
     """Detect PII entities in plain text.
 
@@ -64,7 +65,7 @@ def anonymize_plain_text(
     strategy: AnonymizationStrategy = AnonymizationStrategy.REPLACE,
     entity_strategies: dict[str, AnonymizationStrategy] | None = None,
     language: str = "de",
-    score_threshold: float = 0.35,
+    score_threshold: float = DEFAULT_SCORE_THRESHOLD,
 ) -> tuple[str, list[RecognizerResult]]:
     """Detect and anonymize PII in plain text.
 

src/document_anonymizer/web/routes.py

@@ -19,6 +19,11 @@
 from document_anonymizer.anonymization.engine import anonymize_text
 from document_anonymizer.anonymization.strategies import AnonymizationStrategy
 from document_anonymizer.api.dependencies import get_analyzer, get_anonymizer
+from document_anonymizer.constants import (
+    DEFAULT_SCORE_THRESHOLD,
+    TIER_HIGH_THRESHOLD,
+    TIER_MEDIUM_THRESHOLD,
+)
 from document_anonymizer.document.pdf_handler import (
     IncompleteRedactionError,
     PdfPageLimitExceededError,
@@ -130,9 +135,9 @@ def _template_response(
 
 def _score_to_tier(score: float) -> Tier:
     """Map a confidence score to a review tier."""
-    if score >= 0.7:
+    if score >= TIER_HIGH_THRESHOLD:
         return "high"
-    if score >= 0.5:
+    if score >= TIER_MEDIUM_THRESHOLD:
         return "medium"
     return "low"
 
@@ -363,7 +368,7 @@ def _normalize_line_endings(text: str) -> str:
 async def detect_form(
     request: Request,
     text: Annotated[str, Form(max_length=_MAX_TEXT_LENGTH)] = "",
-    score_threshold: Annotated[float, Form(ge=0.0, le=1.0)] = 0.35,
+    score_threshold: Annotated[float, Form(ge=0.0, le=1.0)] = DEFAULT_SCORE_THRESHOLD,
     file: UploadFile | None = File(default=None),  # noqa: B008
     analyzer: AnalyzerEngine = Depends(get_analyzer),  # noqa: B008
 ) -> HTMLResponse:
@@ -489,7 +494,7 @@ async def anonymize_form(
     request: Request,
     text: Annotated[str, Form(max_length=_MAX_TEXT_LENGTH)] = "",
     strategy: str = Form(default="replace"),
-    score_threshold: Annotated[float, Form(ge=0.0, le=1.0)] = 0.35,
+    score_threshold: Annotated[float, Form(ge=0.0, le=1.0)] = DEFAULT_SCORE_THRESHOLD,
     is_pdf: bool = Form(default=False),
     pdf_b64: str = Form(default=""),
     selected_entities: str = Form(default=""),
@@ -603,7 +608,7 @@ async def anonymize_form(
 async def redact_pdf_form(
     request: Request,
     pdf_b64: str = Form(...),
-    score_threshold: Annotated[float, Form(ge=0.0, le=1.0)] = 0.35,
+    score_threshold: Annotated[float, Form(ge=0.0, le=1.0)] = DEFAULT_SCORE_THRESHOLD,
     selected_entities: str = Form(default=""),
     analyzer: AnalyzerEngine = Depends(get_analyzer),  # noqa: B008
 ) -> Response: