fix(api): expose actual PII confidence scores instead of hardcoded 0.9

Fixes vllm-project#717

Changes:
- Add ClassifyPIIWithDetails() method to classifier that returns full
  entity details including actual confidence scores
- Update DetectPII() service to use ClassifyPIIWithDetails() instead
  of ClassifyPII() to access real confidence values
- Update config.e2e.yaml to use LoRA PII model and add default
  catch-all decision for E2E testing

Before: API returned hardcoded confidence=0.9 for all PII entities
After: API returns actual model confidence scores (0.77-0.99 range)

Test results:
- Email detection: 0.9869 (was 0.9)
- SSN detection: 0.9916, 0.8326 (was 0.9)
- Multiple PII: 0.7734-0.9991 range (was all 0.9)

The underlying LoRA PII model was already producing accurate confidence
scores (fixed in vllm-project#709), but the API layer was discarding them. This
change passes through the actual scores to API consumers.

Signed-off-by: Yossi Ovadia <[email protected]>

Author: yossiovadia

config/testing/config.e2e.yaml

@@ -69,11 +69,11 @@ classifier:
     use_cpu: true
     category_mapping_path: "models/lora_intent_classifier_bert-base-uncased_model/category_mapping.json"
   pii_model:
-    model_id: "models/pii_classifier_modernbert-base_presidio_token_model"  # TODO: Use local model for now before the code can download the entire model from huggingface
-    use_modernbert: true
+    model_id: "models/lora_pii_detector_bert-base-uncased_model"
+    use_modernbert: false  # BERT-based LoRA model (this field is ignored - always auto-detects)
     threshold: 0.7
     use_cpu: true
-    pii_mapping_path: "models/pii_classifier_modernbert-base_presidio_token_model/pii_type_mapping.json"
+    pii_mapping_path: "models/lora_pii_detector_bert-base-uncased_model/pii_type_mapping.json"
 categories:
   - name: business
     description: "Business and management related queries"
@@ -359,6 +359,24 @@ decisions:
           enabled: true
           pii_types_allowed: ["EMAIL_ADDRESS", "PERSON", "GPE", "PHONE_NUMBER", "US_SSN", "CREDIT_CARD"]
 
+  # Default catch-all decision for unmatched requests (E2E PII test fix)
+  # This ensures PII detection is always enabled, even when no specific decision matches
+  - name: "default_decision"
+    description: "Default catch-all decision - blocks all PII for safety"
+    priority: 1  # Lowest priority - only matches if nothing else does
+    rules:
+      operator: "OR"
+      conditions:
+        - type: "always"  # Always matches as fallback
+    modelRefs:
+      - model: "Model-B"
+        use_reasoning: false
+    plugins:
+      - type: "pii"
+        configuration:
+          enabled: true
+          pii_types_allowed: []  # Block ALL PII - empty list means nothing allowed
+
 default_model: "Model-A"
 
 # API Configuration

src/semantic-router/pkg/classification/classifier.go

@@ -884,6 +884,67 @@ func (c *Classifier) ClassifyPIIWithThreshold(text string, threshold float32) ([
 	return result, nil
 }
 
+// ClassifyPIIWithDetails performs PII token classification and returns full entity details including confidence scores
+func (c *Classifier) ClassifyPIIWithDetails(text string) ([]PIIDetection, error) {
+	return c.ClassifyPIIWithDetailsAndThreshold(text, c.Config.PIIModel.Threshold)
+}
+
+// ClassifyPIIWithDetailsAndThreshold performs PII token classification with a custom threshold and returns full entity details
+func (c *Classifier) ClassifyPIIWithDetailsAndThreshold(text string, threshold float32) ([]PIIDetection, error) {
+	if !c.IsPIIEnabled() {
+		return []PIIDetection{}, fmt.Errorf("PII detection is not properly configured")
+	}
+
+	if text == "" {
+		return []PIIDetection{}, nil
+	}
+
+	// Use PII token classifier for entity detection
+	configPath := fmt.Sprintf("%s/config.json", c.Config.PIIModel.ModelID)
+	start := time.Now()
+	tokenResult, err := c.piiInference.ClassifyTokens(text, configPath)
+	metrics.RecordClassifierLatency("pii", time.Since(start).Seconds())
+	if err != nil {
+		return nil, fmt.Errorf("PII token classification error: %w", err)
+	}
+
+	if len(tokenResult.Entities) > 0 {
+		logging.Infof("PII token classification found %d entities", len(tokenResult.Entities))
+	}
+
+	// Convert token entities to PII detections, filtering by threshold
+	var detections []PIIDetection
+	for _, entity := range tokenResult.Entities {
+		if entity.Confidence >= threshold {
+			detection := PIIDetection{
+				EntityType: entity.EntityType,
+				Start:      entity.Start,
+				End:        entity.End,
+				Text:       entity.Text,
+				Confidence: entity.Confidence,
+			}
+			detections = append(detections, detection)
+			logging.Infof("Detected PII entity: %s ('%s') at [%d-%d] with confidence %.3f",
+				entity.EntityType, entity.Text, entity.Start, entity.End, entity.Confidence)
+		}
+	}
+
+	if len(detections) > 0 {
+		// Log unique PII types for compatibility with existing logs
+		uniqueTypes := make(map[string]bool)
+		for _, d := range detections {
+			uniqueTypes[d.EntityType] = true
+		}
+		types := make([]string, 0, len(uniqueTypes))
+		for t := range uniqueTypes {
+			types = append(types, t)
+		}
+		logging.Infof("Detected PII types: %v", types)
+	}
+
+	return detections, nil
+}
+
 // DetectPIIInContent performs PII classification on all provided content
 func (c *Classifier) DetectPIIInContent(allContent []string) []string {
 	var detectedPII []string

src/semantic-router/pkg/services/classification.go

@@ -290,8 +290,8 @@ func (s *ClassificationService) DetectPII(req PIIRequest) (*PIIResponse, error)
 		}, nil
 	}
 
-	// Perform PII detection using the existing classifier
-	piiTypes, err := s.classifier.ClassifyPII(req.Text)
+	// Perform PII detection using the classifier with full details
+	detections, err := s.classifier.ClassifyPIIWithDetails(req.Text)
 	if err != nil {
 		return nil, fmt.Errorf("PII detection failed: %w", err)
 	}
@@ -300,17 +300,19 @@ func (s *ClassificationService) DetectPII(req PIIRequest) (*PIIResponse, error)
 
 	// Build response
 	response := &PIIResponse{
-		HasPII:           len(piiTypes) > 0,
+		HasPII:           len(detections) > 0,
 		Entities:         []PIIEntity{},
 		ProcessingTimeMs: processingTime,
 	}
 
-	// Convert PII types to entities (simplified for now)
-	for _, piiType := range piiTypes {
+	// Convert PII detections to API entities with actual confidence scores
+	for _, detection := range detections {
 		entity := PIIEntity{
-			Type:       piiType,
-			Value:      "[DETECTED]", // Placeholder - would need actual entity extraction
-			Confidence: 0.9,          // Placeholder - would need actual confidence
+			Type:       detection.EntityType,
+			Value:      "[DETECTED]",                  // Redacted for security
+			Confidence: float64(detection.Confidence), // Actual confidence from model
+			StartPos:   detection.Start,
+			EndPos:     detection.End,
 		}
 		response.Entities = append(response.Entities, entity)
 	}