Quick start

pip install mistralai openai pydantic gradio pandas numpy

from pas2 import PAS2

# Initialize the detector
detector = PAS2(
    mistral_api_key="your_mistral_key",  # Optional if set in environment
    openai_api_key="your_openai_key"     # Optional if set in environment
)

# Run hallucination detection
query = "Who was the first person to land on the moon?"
results = detector.detect_hallucination(query, n_paraphrases=3)

# Check the results
if results["hallucination_detected"]:
    print(f"⚠️  Hallucination detected with {results['confidence_score']:.0%} confidence")
    print(f"Summary: {results['summary']}")
    print(f"\nConflicting facts found: {len(results['conflicting_facts'])}")
else:
    print(f"✓ No hallucination detected ({results['confidence_score']:.0%} confidence)")
    print(f"Summary: {results['summary']}")

{
    "original_query": "Who was the first person to land on the moon?",
    "original_response": "Neil Armstrong was the first person...",
    "paraphrased_queries": [
        "Who was the initial individual to set foot on the moon?",
        "Which person first landed on the lunar surface?",
        "Who made the first moon landing?"
    ],
    "paraphrased_responses": [
        "Neil Armstrong became the first human...",
        "The first person to land on the moon was Neil Armstrong...",
        "Neil Armstrong accomplished the first moon landing..."
    ],
    "hallucination_detected": False,
    "confidence_score": 0.95,
    "conflicting_facts": [],
    "reasoning": "All responses consistently identify Neil Armstrong...",
    "summary": "No factual inconsistencies found across responses"
}

def progress_callback(stage, **kwargs):
    if stage == "generating_paraphrases":
        print("Generating paraphrases...")
    elif stage == "responses_progress":
        completed = kwargs.get("completed_responses", 0)
        total = kwargs.get("total_responses", 0)
        print(f"Getting responses: {completed}/{total}")
    elif stage == "judging":
        print("Analyzing for hallucinations...")
    elif stage == "complete":
        print("Detection complete!")

detector = PAS2(progress_callback=progress_callback)
results = detector.detect_hallucination("What is the capital of France?")

# More paraphrases = more thorough detection but slower
results = detector.detect_hallucination(
    "What year did World War II end?",
    n_paraphrases=5  # Generate 5 paraphrases
)

# From pas2.py lines 302-366
def judge_hallucination(self, 
                       original_query: str, 
                       original_response: str, 
                       paraphrased_queries: List[str], 
                       paraphrased_responses: List[str]) -> HallucinationJudgment:
    """
    Use OpenAI's o3-mini as a judge to detect hallucinations in the responses
    """
    # The judge analyzes all responses for factual inconsistencies
    response = self.openai_client.chat.completions.create(
        model=self.openai_model,  # "o3-mini"
        messages=[...],
        response_format={"type": "json_object"}
    )

class HallucinationJudgment(BaseModel):
    hallucination_detected: bool
    confidence_score: float  # Between 0-1
    conflicting_facts: List[Dict[str, Any]]
    reasoning: str
    summary: str

try:
    detector = PAS2(
        mistral_api_key=os.environ.get("MISTRAL_API_KEY"),
        openai_api_key=os.environ.get("OPENAI_API_KEY")
    )
    results = detector.detect_hallucination(query)
except ValueError as e:
    print(f"Configuration error: {e}")
except Exception as e:
    print(f"Detection failed: {e}")