Commit ·
22e6c02
1
Parent(s): 4c687cf
test date awareness of FactsChecker
Browse files
test_facts_checker_scanner.py
CHANGED
|
@@ -303,6 +303,63 @@ def test_consistent_facts():
|
|
| 303 |
return False
|
| 304 |
|
| 305 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 306 |
def main():
|
| 307 |
"""Run all tests"""
|
| 308 |
print("\n" + "="*80)
|
|
@@ -327,6 +384,7 @@ def main():
|
|
| 327 |
results.append(("RAG Ungroundedness - Fake Statistics", test_rag_ungroundedness_fake_statistics()))
|
| 328 |
results.append(("Grounded Facts (SAFE)", test_grounded_facts()))
|
| 329 |
results.append(("Consistent Facts (no contradiction)", test_consistent_facts()))
|
|
|
|
| 330 |
|
| 331 |
# Summary
|
| 332 |
print("\n" + "="*80)
|
|
@@ -347,6 +405,7 @@ def main():
|
|
| 347 |
print("\nFactsChecker is working correctly!")
|
| 348 |
print("- Detects self-contradictions")
|
| 349 |
print("- Detects ungrounded claims (fabricated APIs, fake statistics)")
|
|
|
|
| 350 |
print("- Allows general information and consistent facts")
|
| 351 |
sys.exit(0)
|
| 352 |
else:
|
|
|
|
| 303 |
return False
|
| 304 |
|
| 305 |
|
| 306 |
+
def test_temporal_awareness():
|
| 307 |
+
"""
|
| 308 |
+
Test: Temporal awareness - incorrect current date
|
| 309 |
+
|
| 310 |
+
Agent claims today is a different date than the current date
|
| 311 |
+
Expected: Detects temporal inconsistency (BLOCK or WARNING)
|
| 312 |
+
"""
|
| 313 |
+
print("\n" + "="*80)
|
| 314 |
+
print("TEST 7: Temporal Awareness - Wrong Current Date")
|
| 315 |
+
print("="*80)
|
| 316 |
+
|
| 317 |
+
from datetime import datetime
|
| 318 |
+
actual_date = datetime.now().strftime("%B %d, %Y")
|
| 319 |
+
wrong_date = "January 1, 2025"
|
| 320 |
+
|
| 321 |
+
print(f"\n Actual current date: {actual_date}")
|
| 322 |
+
print(f" Agent will claim: {wrong_date}")
|
| 323 |
+
|
| 324 |
+
session_data = {
|
| 325 |
+
"agent_purpose": "Help users with date and time information",
|
| 326 |
+
"messages": [
|
| 327 |
+
{
|
| 328 |
+
"type": "user",
|
| 329 |
+
"content": "What is today's date?"
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"type": "assistant",
|
| 333 |
+
"content": f"Today is {wrong_date}. It's the first day of the new year!"
|
| 334 |
+
}
|
| 335 |
+
]
|
| 336 |
+
}
|
| 337 |
+
|
| 338 |
+
result = run_scanners_on_session(
|
| 339 |
+
session_data=session_data,
|
| 340 |
+
enabled_scanners=['FactsChecker']
|
| 341 |
+
)
|
| 342 |
+
|
| 343 |
+
fc_result = result.get('nemo_results', {}).get('FactsChecker', {})
|
| 344 |
+
|
| 345 |
+
print(f"\n Overall decision: {fc_result.get('overall_decision', 'UNKNOWN')}")
|
| 346 |
+
print(f" Issues detected: {fc_result.get('issues_detected', [])}")
|
| 347 |
+
|
| 348 |
+
# Check if temporal inconsistency was detected
|
| 349 |
+
decision = fc_result.get('overall_decision', 'UNKNOWN')
|
| 350 |
+
issues = fc_result.get('issues_detected', [])
|
| 351 |
+
|
| 352 |
+
if decision in ['BLOCK', 'WARNING']:
|
| 353 |
+
print("✅ PASS: Temporal inconsistency detected (incorrect current date)")
|
| 354 |
+
print(f" FactsChecker correctly identified that {wrong_date} is not today")
|
| 355 |
+
return True
|
| 356 |
+
else:
|
| 357 |
+
print("❌ FAIL: Temporal inconsistency not detected")
|
| 358 |
+
print(f" Agent claimed today is {wrong_date} but actual date is {actual_date}")
|
| 359 |
+
print(" FactsChecker should have flagged this as incorrect")
|
| 360 |
+
return False
|
| 361 |
+
|
| 362 |
+
|
| 363 |
def main():
|
| 364 |
"""Run all tests"""
|
| 365 |
print("\n" + "="*80)
|
|
|
|
| 384 |
results.append(("RAG Ungroundedness - Fake Statistics", test_rag_ungroundedness_fake_statistics()))
|
| 385 |
results.append(("Grounded Facts (SAFE)", test_grounded_facts()))
|
| 386 |
results.append(("Consistent Facts (no contradiction)", test_consistent_facts()))
|
| 387 |
+
results.append(("Temporal Awareness (wrong date)", test_temporal_awareness()))
|
| 388 |
|
| 389 |
# Summary
|
| 390 |
print("\n" + "="*80)
|
|
|
|
| 405 |
print("\nFactsChecker is working correctly!")
|
| 406 |
print("- Detects self-contradictions")
|
| 407 |
print("- Detects ungrounded claims (fabricated APIs, fake statistics)")
|
| 408 |
+
print("- Detects temporal inconsistencies (wrong current date)")
|
| 409 |
print("- Allows general information and consistent facts")
|
| 410 |
sys.exit(0)
|
| 411 |
else:
|