Quick Start Guide
This guide will help you get AuraGen up and running in just a few minutes.
Prerequisites
Before starting, ensure you have:
Python 3.8+ installed
AuraGen installed (see Installation)
At least one API key (OpenAI or DeepInfra recommended)
Step 1: Configure API Keys
AuraGen supports multiple API providers. Start by configuring at least one:
python config/configure_api_keys.py
This interactive tool will guide you through:
Selecting API Key Type: Choose from existing types or add custom ones
Entering API Key: Securely input your API key (hidden input)
Choosing Storage: Save to project
.envfile or system environment
Note
We recommend using the project .env file for easy project-specific configuration.
Example session:
┌────────────────────────────────────────┐
│ Setup │
└────────────────────────────────────────┘
Current API Key Values (masked)
┏━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┓
┃ Key Type ┃ Env Var ┃ Value ┃
┡━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━┩
│ openai_api_key │ OPENAI_API_KEY │ <not set> │
│ deepinfra_api_key│ DEEPINFRA_API_KEY│ <not set> │
└────────────────────┴──────────────────┴──────────────────┘
Step 2: Basic Configuration
Configure the generation settings in config/generation.yaml:
# Basic configuration
generation:
batch_size: 10
externalAPI_generation: false # Use OpenAI (true for external APIs)
# OpenAI settings
openai:
api_key_type: "openai_api_key"
model: "gpt-4o"
temperature: 1.0
max_tokens: 2048
Important
The externalAPI_generation setting determines which API service to use:
false: Use OpenAI APItrue: Use external API (DeepInfra, etc.)
Step 3: Generate Your First Dataset
Run the complete generation and injection pipeline:
python generate_and_inject.py
This command will:
Load Scenarios: Read all scenarios from
config/AgentSafetyBench/Generate Harmless Trajectories: Create clean agent interactions
Apply Risk Injection: Introduce realistic risks while maintaining plausibility
Save Results: Output files to
generated_records/
Expected output:
🚀 Starting AuraGen Pipeline...
📊 Loaded 150 scenarios
⚙️ Using OpenAI API (gpt-4o)
🔄 Generating harmless trajectories...
✅ Generated 1,500 harmless records
💉 Injecting risks...
✅ Created 1,500 risky trajectories
💾 Saved to generated_records/
🎉 Pipeline completed successfully!
Step 4: Examine the Results
The generated files will be saved in the generated_records/ directory:
generated_records/
├── all_scenarios_openai_20241215_143022.json # Harmless trajectories
└── all_injected_openai_20241215_143022.json # Risk-injected trajectories
Each record contains:
{
"scenario_name": "email_assistant",
"user_request": "Help me write a professional email",
"agent_action": "draft_email",
"agent_response": "I'll help you create a professional email...",
"metadata": {
"timestamp": 1703172602,
"api_model": "gpt-4o",
"risk_type": "privacy_breach",
"scenario_metadata": {
"industry": "healthcare",
"urgency_level": "medium"
}
}
}
Understanding the Data Structure
Harmless Trajectories
These represent clean, appropriate agent behavior:
user_request: The input from the user
agent_action: The action the agent takes
agent_response: The agent’s response to the user
metadata: Contextual information and constraints
Risk-Injected Trajectories
These contain introduced risks while maintaining plausibility:
Same structure as harmless trajectories
Additional
risk_typein metadataModified
agent_actionoragent_responsewith realistic risks
Common Risk Types
privacy_breach: Unauthorized access to personal informationmisinformation: Spreading false or misleading informationbias_amplification: Reinforcing harmful stereotypesunauthorized_action: Actions beyond the agent’s scopeavailability_disruption: Service interruptions or failures
Next Steps
Now that you have AuraGen running, explore these advanced features:
Configuration - Detailed configuration options
Scenarios - Understanding and customizing scenarios
Risk Injection - Advanced risk injection techniques
advanced/custom_scenarios - Creating your own scenarios
Common Issues
“Environment variable not set” Error
Make sure you’ve configured your API keys:
python config/configure_api_keys.py
Empty or Failed Generation
Check your API key validity and internet connection. Also verify the model name in your configuration.
Permission Errors
Ensure you have write permissions in the project directory:
chmod -R 755 /path/to/agentic-guardian
Need Help?
Check the advanced/troubleshooting guide
Review the full Configuration documentation
Visit our GitHub repository for issues and discussions