Skip to main content
PAS2 uses environment variables to manage API keys and configuration settings. This guide covers all available configuration options.

​
Environment variables

PAS2 supports environment variables for API authentication and configuration. Create a .env file in your project root:
.env
HF_MISTRAL_API_KEY=your_mistral_api_key_here
HF_OPENAI_API_KEY=your_openai_api_key_here
The HF_ prefix is prioritized for Hugging Face Spaces deployment. The system also supports MISTRAL_API_KEY and OPENAI_API_KEY as fallback options.

​
Configuration priority

PAS2 checks for API keys in the following order:
1

Direct parameter

API keys passed directly to the PAS2 class constructor:
pas2 = PAS2(mistral_api_key="key", openai_api_key="key")
2

Hugging Face environment variables

Environment variables with HF_ prefix:
  • HF_MISTRAL_API_KEY
  • HF_OPENAI_API_KEY
3

Standard environment variables

Standard environment variable names:
  • MISTRAL_API_KEY
  • OPENAI_API_KEY

​
Model configuration

PAS2 uses specific models for different tasks:

​
Default models

  • Mistral model: mistral-large-latest - Used for paraphrase generation and response generation
  • OpenAI model: o3-mini - Used as the judge model for hallucination detection
Changing the default models may affect hallucination detection accuracy. The system is optimized for the models listed above.

​
Paraphrase settings

Configure the number of paraphrases generated for each query: 
# Default: 3 paraphrases
results = pas2.detect_hallucination(query, n_paraphrases=3)

# Increase for more thorough analysis (slower)
results = pas2.detect_hallucination(query, n_paraphrases=5)

# Reduce for faster processing (less comprehensive)
results = pas2.detect_hallucination(query, n_paraphrases=2)
More paraphrases increase accuracy but also increase API costs and processing time.

​
Database configuration

PAS2 uses SQLite for persistent feedback storage:
  • Default path: /data/feedback.db (Hugging Face Spaces)
  • Fallback path: ./temp_data/feedback.db (local development)
The database is automatically initialized on first run and stores:
  • Original queries and responses
  • Paraphrased queries and responses
  • Hallucination detection results
  • User feedback
  • Confidence scores and reasoning

​
Logging configuration

PAS2 includes built-in logging for debugging and monitoring: 
import logging

# Default level: INFO
logging.basicConfig(level=logging.INFO)

# For more detailed logs:
logging.basicConfig(level=logging.DEBUG)

# For production (errors only):
logging.basicConfig(level=logging.ERROR)

​
Performance tuning

​
Parallel processing

PAS2 uses ThreadPoolExecutor for parallel API calls with a default maximum of 5 workers: 
max_workers = min(len(queries), 5)
This balances speed with API rate limits.

​
Progress tracking

Enable progress callbacks for real-time updates: 
def progress_callback(stage, **kwargs):
    print(f"Stage: {stage}, Details: {kwargs}")

pas2 = PAS2(
    mistral_api_key="key",
    openai_api_key="key",
    progress_callback=progress_callback
)

​
Next steps

Build docs developers (and LLMs) love

Get started for freeTalk to us