Skip to content

Commit 980405f

Browse files
yossiovadiayossiovadia
authored
Feature/improve pii extproc testing (#335)
* feat: enhance PII detection testing with comprehensive ExtProc validation 🔍 ENHANCED PII TESTING FRAMEWORK: - Added comprehensive ExtProc PII detection test (TEST 3.5) - Tests differential behavior between PII and safe content - Validates production pipeline PII handling capabilities - Monitors routing decisions, processing times, and blocking behavior 📋 IMPROVED TEST COVERAGE: - Enhanced test case generation with cache-busting timestamps - Added comprehensive PII pattern analysis across multiple entity types - Better detection of ExtProc PII filtering mechanisms - More detailed logging and result analysis ⚙️ SMART PII POLICY CONFIGURATION: - Model-A: Strict PII policy (allow_by_default: false, EMAIL_ADDRESS only) - Model-B: Permissive PII policy (allow_by_default: true, all PII types) - Mixed policy approach enables better testing of PII routing behavior 📊 TEST CAPABILITIES: - Detects PII blocking vs routing-only behavior - Monitors differential model selection based on PII content - Validates security policy enforcement in production pipeline - Comprehensive analysis of ExtProc PII detection indicators This establishes a comprehensive testing framework that will reveal any gaps in PII detection and policy enforcement across the entire semantic router pipeline. Signed-off-by: Yossi Ovadia <[email protected]> * style: apply black formatting to PII detection test Apply automatic Python code formatting from black to ensure consistent code style across the test file. No functional changes - only formatting improvements including: - Trailing commas for better diffs - Line wrapping for readability - Consistent spacing around operators Signed-off-by: Yossi Ovadia <[email protected]> --------- Signed-off-by: Yossi Ovadia <[email protected]>
1 parent 0270850 commit 980405f

File tree

2 files changed

+213
-4
lines changed

2 files changed

+213
-4
lines changed

‎config/config.e2e.yaml‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -61,14 +61,14 @@ model_config:
6161
reasoning_family: "qwen3" # This model uses Qwen reasoning syntax
6262
preferred_endpoints: ["qwen-endpoint"]
6363
pii_policy:
64-
allow_by_default: true
65-
pii_types_allowed: ["EMAIL_ADDRESS", "PERSON", "GPE", "PHONE_NUMBER"]
64+
allow_by_default: false # Strict PII blocking model
65+
pii_types_allowed: ["EMAIL_ADDRESS"] # Only allow emails
6666
"Model-B":
6767
use_reasoning: false
6868
preferred_endpoints: ["tinyllama-endpoint"]
6969
pii_policy:
70-
allow_by_default: true
71-
pii_types_allowed: ["EMAIL_ADDRESS", "PERSON", "GPE", "PHONE_NUMBER"]
70+
allow_by_default: true # Permissive PII model for safe routing
71+
pii_types_allowed: ["EMAIL_ADDRESS", "PERSON", "GPE", "PHONE_NUMBER", "US_SSN", "CREDIT_CARD"]
7272

7373
# Classifier configuration for text classification
7474
classifier:

‎e2e-tests/06-pii-detection-test.py‎

Lines changed: 209 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -23,6 +23,7 @@
2323

2424
import json
2525
import sys
26+
import time
2627
import unittest
2728
import uuid
2829

@@ -467,6 +468,214 @@ def test_extproc_pii_filtering(self):
467468
message="✅ Request blocked: ExtProc may have PII filtering enabled",
468469
)
469470

471+
def test_extproc_pii_detection_comprehensive(self):
472+
"""
473+
TEST 3.5: ExtProc PII Detection - Comprehensive Production Pipeline Test
474+
475+
WHAT IS TESTED: Full production pipeline PII detection capabilities
476+
PIPELINE: client → envoy → extProc → semantic-router → model routing
477+
EXPECTED RESULT: ⚠️ VALIDATION - Tests actual PII detection in production flow
478+
479+
This tests the MAIN USE CASE: whether ExtProc actually detects PII and enforces policies
480+
in the production request pipeline, not just whether responses leak PII.
481+
"""
482+
self.print_test_header(
483+
"TEST 3.5: ExtProc PII Detection - Comprehensive Production Pipeline",
484+
"Expected: ⚠️ VALIDATION - Tests if ExtProc actually detects PII in production flow",
485+
)
486+
487+
# Generate unique PII test cases
488+
pii_cases, safe_cases = generate_unique_test_cases()
489+
490+
# Test 1: High-value PII that should trigger policy enforcement
491+
high_value_pii = pii_cases[0]["text"] # SSN case
492+
493+
# Test 2: Safe content that should pass through
494+
safe_content = safe_cases[0]["text"] # Math problem
495+
496+
test_cases = [
497+
{
498+
"name": "High-Value PII Content",
499+
"content": high_value_pii,
500+
"expected_pii": True,
501+
"expected_action": "May block or route to PII-safe model",
502+
},
503+
{
504+
"name": "Safe Content",
505+
"content": safe_content,
506+
"expected_pii": False,
507+
"expected_action": "Should route normally",
508+
},
509+
]
510+
511+
results = []
512+
513+
for test_case in test_cases:
514+
print(f"\n🔍 Testing: {test_case['name']}")
515+
print(f"Content: {test_case['content'][:60]}...")
516+
517+
# Send request through full ExtProc pipeline
518+
payload = {
519+
"model": "auto",
520+
"messages": [{"role": "user", "content": test_case["content"]}],
521+
"temperature": 0.1,
522+
}
523+
524+
# Monitor for specific behaviors that indicate PII detection
525+
start_time = time.time()
526+
527+
response = requests.post(
528+
f"{ENVOY_URL}{OPENAI_ENDPOINT}",
529+
headers={"Content-Type": "application/json"},
530+
json=payload,
531+
timeout=60,
532+
)
533+
534+
processing_time = (time.time() - start_time) * 1000
535+
536+
# Analyze response for PII detection indicators
537+
request_allowed = response.status_code == 200
538+
539+
# Check for PII policy violation response
540+
is_pii_blocked = response.status_code == 403 or response.status_code == 400
541+
542+
# Analyze response content
543+
response_content = ""
544+
model_used = "unknown"
545+
if request_allowed:
546+
try:
547+
response_json = response.json()
548+
if "choices" in response_json:
549+
response_content = response_json["choices"][0]["message"][
550+
"content"
551+
]
552+
model_used = response_json.get("model", "unknown")
553+
except:
554+
response_content = "Could not parse response"
555+
556+
# Check for PII-related routing decisions
557+
# ExtProc may route PII content to specific models or add safety measures
558+
pii_safe_routing = model_used in [
559+
"Model-A",
560+
"safe-model",
561+
"pii-compliant-model",
562+
]
563+
564+
# Check response quality/safety indicators
565+
response_looks_safe = (
566+
len(response_content) > 50
567+
and "sorry" not in response_content.lower()
568+
and "can't help" not in response_content.lower()
569+
)
570+
571+
# Analyze processing time (PII detection may add latency)
572+
baseline_time = 1000 # ms
573+
has_extra_processing = processing_time > baseline_time * 1.5
574+
575+
result = {
576+
"test_case": test_case["name"],
577+
"content": test_case["content"][:50] + "...",
578+
"expected_pii": test_case["expected_pii"],
579+
"request_allowed": request_allowed,
580+
"is_blocked": is_pii_blocked,
581+
"model_used": model_used,
582+
"processing_time_ms": processing_time,
583+
"has_extra_processing": has_extra_processing,
584+
"response_safe": response_looks_safe,
585+
"response_preview": (
586+
response_content[:100] + "..." if response_content else "No content"
587+
),
588+
}
589+
590+
results.append(result)
591+
592+
print(f" 📊 Results:")
593+
print(f" Request Status: {response.status_code}")
594+
print(f" Model Used: {model_used}")
595+
print(f" Processing Time: {processing_time:.1f}ms")
596+
print(f" Expected PII: {test_case['expected_pii']}")
597+
print(f" Blocked: {is_pii_blocked}")
598+
print(f" Extra Processing: {has_extra_processing}")
599+
600+
# Analyze results for PII detection patterns
601+
pii_test = results[0] # High-value PII test
602+
safe_test = results[1] # Safe content test
603+
604+
# Check for differential behavior between PII and safe content
605+
differential_routing = pii_test["model_used"] != safe_test["model_used"]
606+
differential_processing = (
607+
abs(pii_test["processing_time_ms"] - safe_test["processing_time_ms"]) > 200
608+
)
609+
differential_blocking = pii_test["is_blocked"] != safe_test["is_blocked"]
610+
611+
# Overall PII detection indicators
612+
pii_detection_indicators = []
613+
if pii_test["is_blocked"]:
614+
pii_detection_indicators.append("Request blocking")
615+
if differential_routing:
616+
pii_detection_indicators.append("Model routing differences")
617+
if differential_processing:
618+
pii_detection_indicators.append("Processing time differences")
619+
if not pii_test["response_safe"] and safe_test["response_safe"]:
620+
pii_detection_indicators.append("Response quality differences")
621+
622+
# Final assessment
623+
pii_detection_evidence = len(pii_detection_indicators) > 0
624+
625+
self.print_response_info(
626+
response, # Use last response for HTTP details
627+
{
628+
"Test Cases": len(test_cases),
629+
"PII Detection Evidence": (
630+
"✅ YES" if pii_detection_evidence else "❌ NO"
631+
),
632+
"Detection Indicators": (
633+
", ".join(pii_detection_indicators)
634+
if pii_detection_indicators
635+
else "None found"
636+
),
637+
"PII Content Model": pii_test["model_used"],
638+
"Safe Content Model": safe_test["model_used"],
639+
"Differential Routing": "✅ YES" if differential_routing else "❌ NO",
640+
"PII Request Blocked": "✅ YES" if pii_test["is_blocked"] else "❌ NO",
641+
"Overall Assessment": (
642+
"✅ PII DETECTION ACTIVE"
643+
if pii_detection_evidence
644+
else "⚠️ NO CLEAR PII DETECTION"
645+
),
646+
},
647+
)
648+
649+
# Print detailed analysis
650+
print(f"\n📋 Detailed ExtProc PII Analysis:")
651+
for result in results:
652+
status = (
653+
"🔒"
654+
if result["is_blocked"]
655+
else "✅" if result["request_allowed"] else "❌"
656+
)
657+
print(f" {status} {result['test_case']}")
658+
print(f" Content: {result['content']}")
659+
print(
660+
f" Model: {result['model_used']}, Time: {result['processing_time_ms']:.1f}ms"
661+
)
662+
print(f" Status: {'Blocked' if result['is_blocked'] else 'Allowed'}")
663+
664+
if pii_detection_evidence:
665+
self.print_test_result(
666+
passed=True,
667+
message=f"✅ ExtProc PII detection evidence found: {', '.join(pii_detection_indicators)}",
668+
)
669+
else:
670+
self.print_test_result(
671+
passed=False,
672+
message="⚠️ No clear evidence of ExtProc PII detection in production pipeline",
673+
)
674+
print(
675+
"📝 NOTE: This may indicate PII detection is not active in ExtProc or"
676+
)
677+
print(" PII policies are configured to allow all content through")
678+
470679
def test_multiple_pii_types_analysis(self):
471680
"""
472681
TEST 4: Multiple PII Types Pattern Analysis

0 commit comments

Comments
 (0)