stinger-guardrails-alpha 0.1.0a4

AI Guardrails Framework - Comprehensive LLM safety toolkit (Alpha)

🚀 Stinger - AI Guardrails Framework

A powerful, easy-to-use Python framework for safeguarding LLM applications with comprehensive content filtering and moderation capabilities.

Alpha Release Available! Install with pip install stinger-guardrails-alpha

✨ Features

• 🛡️ Comprehensive Guardrails: Toxicity detection, PII protection, code generation prevention, and more
• 🔒 Security Audit Trail: Complete logging of all security decisions for compliance and forensics
• 🎯 Simple API: Get started in 3 lines of code
• ⚡ High Performance: Async-ready with synchronous convenience wrapper
• 🔧 Configurable: YAML-based configuration with runtime updates
• 🧪 Production Ready: Comprehensive testing and error handling
• 📚 Well Documented: Complete API reference and examples

🚀 Quick Start

Installation

# Install the alpha release
pip install stinger-guardrails-alpha

# Or install from source for development
pip install .

Basic Usage

from stinger import GuardrailPipeline
from stinger.core import audit

# Enable security audit trail (zero-config)
audit.enable()  # Tracks all security decisions

# Create a pipeline from preset
pipeline = GuardrailPipeline.from_preset("customer_service")

# Check input content
result = pipeline.check_input("My credit card is 4532-1234-5678-9012")
if result['blocked']:
    print(f"Input blocked: {result['reasons']}")
    # Audit trail automatically logs: user input, guardrail decision, reasons

# Check output content
result = pipeline.check_output("Here's the code: import os; os.system('rm -rf /')")
if result['blocked']:
    print(f"Output blocked: {result['reasons']}")

🖥️ Command Line Interface (CLI)

After installing Stinger, you can use the CLI:

stinger demo
stinger check-prompt "My SSN is 123-45-6789."
stinger check-response "Here is your password: hunter2"

🌟 Interactive Demos

Web Demo - Real-Time Guardrail Visualization

Experience Stinger's power through our interactive web interface that shows guardrails in action:

# Start the web demo (simplest method)
cd demos/web_demo
python start_demo.py

# Or run in background mode (useful for tools with timeouts)
python start_demo.py --detached

# Open http://127.0.0.1:8000 in your browser (use HTTP, not HTTPS)

Features:

• 💬 Interactive chat interface with real-time guardrail feedback
• 🚦 Visual indicators showing which guardrails triggered
• 📊 Live audit trail visualization
• 🔄 Switch between presets (customer service, medical, financial)
• ⚡ See guardrails activate as you type

Management Console - System Monitoring Dashboard

Monitor your Stinger deployment with our real-time management console:

# Start the management console (simplest method)
cd management-console
./start_console.sh

# Or run in background mode (useful for tools with timeouts)
./start_console.sh --detached

# Open http://localhost:3001 in your browser

Features:

• 📈 Real-time metrics and performance monitoring
• 🔍 Active conversation tracking
• 📊 Guardrail trigger statistics
• 🏥 System health monitoring
• 📉 Historical data visualization

🛡️ Available Guardrails

Stinger offers multiple levels of protection with both fast regex-based and sophisticated AI-powered guardrails:

🚀 Simple/Fast Guardrails (No API Key Required)

• Simple PII Detection: Regex-based detection of SSNs, credit cards, emails, phone numbers
• Simple Toxicity Detection: Keyword-based profanity and hate speech filtering
• Simple Code Generation: Pattern-based code snippet detection
• Keyword Blocking: Block specific words or phrases
• URL Filtering: Block or allow specific domains
• Length Limiting: Control input/output length
• Regex Filtering: Custom pattern matching

🤖 AI-Powered Guardrails (Requires OpenAI API Key)

• AI PII Detection: Context-aware PII detection using language models
• AI Toxicity Detection: Nuanced understanding of harmful content
• AI Code Generation: Sophisticated code pattern recognition
• Content Moderation: General inappropriate content detection
• Topic Filtering: AI-based allowed/blocked topic enforcement

🔐 Prompt Injection Protection (Three Levels)

1. Quick/Local Detection: Fast pattern matching for common injection attempts
2. AI-Powered Detection: Single-turn analysis using language models
3. Conversation-Aware AI: Multi-turn context analysis for sophisticated attacks
   • Configurable strategies: 'recent', 'suspicious', or 'mixed' context
   • Risk levels: low, medium, high, critical

🎯 Usage Examples

# Use simple guardrails for speed (no API key needed)
pipeline = GuardrailPipeline.from_preset("customer_service")  # Uses simple versions

# Enable AI guardrails for better accuracy (requires API key)
export OPENAI_API_KEY="sk-..."
pipeline = GuardrailPipeline.from_preset("medical")  # Uses AI versions

# Mix and match as needed
config = {
    "input": [
        {"type": "simple_pii_detection"},     # Fast PII check
        {"type": "ai_toxicity_detection"},    # AI toxicity check
        {"type": "prompt_injection",          # Multi-turn injection detection
         "config": {"conversation_aware": True}}
    ]
}

🔒 Security Audit Trail

Stinger provides comprehensive security audit logging for compliance and forensic analysis:

Zero-Config Audit Trail

from stinger.core import audit

# Enable with smart defaults (just works!)
audit.enable()

# Or specify destination
audit.enable("./logs/security.log")

# With PII redaction for compliance
audit.enable("./logs/audit.log", redact_pii=True)

Complete Security Tracking

• User Prompts: All user inputs logged with attribution
• LLM Responses: All AI responses tracked with context
• Guardrail Decisions: Every security decision with full reasoning
• Conversation Flow: Complete conversation reconstruction
• User Attribution: IP, session, user ID tracking for forensics

Compliance-Standard Ready

• PII Redaction Capabilities: Automatically redact sensitive data while preserving audit value (useful for GDPR, HIPAA, and other privacy regulations)
• Complete Audit Trail: Comprehensive logging suitable for enterprise security reviews
• Data Retention Controls: Configurable retention policies for different data types
• Forensic Analysis: Full incident reconstruction capabilities
• Export Formats: Generate compliance-ready reports in standard formats

Audit Trail Features

• Smart Environment Detection: Auto-configures for dev/prod/docker
• Async Buffering: Background processing with <10ms latency impact
• Query Tools: Easy audit trail searching and analysis
• Export Capabilities: CSV/JSON export for compliance reporting
• Cannot be disabled in production: Ensures audit trail integrity

📋 Configuration

Stinger uses YAML configuration files:

version: "1.0"

pipeline:
  input:
    - name: toxicity_check
      type: simple_toxicity_detection
      enabled: true
      confidence_threshold: 0.7
      categories: [hate_speech, harassment, threats]
    
    - name: pii_check
      type: simple_pii_detection
      enabled: true
      confidence_threshold: 0.8
      categories: [credit_card, ssn, email]
  
  output:
    - name: code_generation_check
      type: simple_code_generation
      enabled: true
      confidence_threshold: 0.6
      categories: [programming_keywords, code_blocks]

🎯 API Reference

GuardrailPipeline

The main class for using Stinger guardrails.

# Initialize
pipeline = GuardrailPipeline("config.yaml")

# Check content
result = pipeline.check_input(content)
result = pipeline.check_output(content)

# Get status
status = pipeline.get_guardrail_status()

# Dynamic configuration
pipeline.enable_guardrail("toxicity_check")
pipeline.disable_guardrail("pii_check")
pipeline.update_guardrail_config("toxicity_check", {"confidence_threshold": 0.9})

Result Format

{
    'blocked': bool,           # Whether content was blocked
    'warnings': List[str],     # List of warning messages
    'reasons': List[str],      # List of blocking reasons
    'details': Dict[str, Any], # Detailed results from each guardrail
    'pipeline_type': str       # Type of pipeline ("input" or "output")
}

Audit Trail API

from stinger.core import audit

# Enable audit trail
audit.enable("./logs/audit.log")

# Query audit trail
records = audit.query(
    conversation_id="conv_123",
    user_id="user_456",
    last_hour=True
)

# Export for compliance
audit.export_csv("./logs/audit.log", "compliance_report.csv")

# Get performance stats
stats = audit.get_stats()
print(f"Queued: {stats['queued']}, Written: {stats['written']}")

📖 Examples

Basic Usage

from stinger import GuardrailPipeline

pipeline = GuardrailPipeline("config.yaml")

# Simple content checking
result = pipeline.check_input("User input here")
if result['blocked']:
    print(f"Blocked: {result['reasons']}")
elif result['warnings']:
    print(f"Warnings: {result['warnings']}")
else:
    print("Content approved")

Advanced Usage

from stinger import GuardrailPipeline, audit

# Enable security audit trail
audit.enable("./logs/audit.log", redact_pii=True)

# Initialize with custom config
pipeline = GuardrailPipeline("my_config.yaml")

# Get pipeline status
status = pipeline.get_guardrail_status()
print(f"Pipeline has {status['total_enabled']} enabled guardrails")

# Dynamically configure guardrails
pipeline.disable_guardrail("pii_check")
pipeline.enable_guardrail("toxicity_check")

# Update configuration
pipeline.update_guardrail_config("toxicity_check", {
    'confidence_threshold': 0.9
})

# Process content with detailed results
result = pipeline.check_input("Test content")
print(f"Blocked: {result['blocked']}")
print(f"Reasons: {result['reasons']}")
print(f"Warnings: {result['warnings']}")
print(f"Details: {result['details']}")

# All security decisions automatically logged to audit trail

🧪 Testing

Run the demo to see Stinger in action:

# Run the tech support demo
cd demos/tech_support
python3 demo.py

# Run the simple example
python3 examples/getting_started/01_basic_installation.py

📚 Learning Resources

Examples (/examples)

Start here - Minimal, focused code examples that mirror the Getting Started guide:

• 01_basic_installation.py - Installation and basic setup
• 02_simple_guardrail.py - Simple guardrail usage
• 03_global_rate_limiting.py - Rate limiting configuration
• 04_conversation_api.py - Conversation-based filtering
• 05_conversation_rate_limiting.py - Conversation rate limiting
• 06_health_monitoring.py - Health monitoring and status
• 07_cli_and_yaml_config.py - CLI and YAML configuration
• 08_security_audit_trail.py - Security audit trail setup and usage
• 09_troubleshooting_and_testing.py - Testing and debugging

Run examples:

cd examples/getting_started
python 01_basic_installation.py

Demos (/demos)

Advanced demonstrations showcasing specific features and scenarios:

• conversation_aware_prompt_injection_demo.py - Advanced prompt injection detection
• global_rate_limiting_demo.py - Rate limiting demonstrations
• topic_filter_demo.py - Topic-based filtering
• tech_support/ - Complete tech support scenario with audit trail

Run demos:

cd demos
python conversation_aware_prompt_injection_demo.py

Learning Path

1. Start with examples - Run through the numbered examples in order
2. Explore demos - Try the advanced demonstrations
3. Check the tech support scenario - See a complete real-world implementation
4. Review API docs - Deep dive into the complete API reference

🔧 Development

Note: Stinger uses a modern src/ layout. All package code is under src/stinger/.

Installation from Source

git clone https://github.com/virtualsteve-star/stinger.git
cd stinger
pip install -e .

Running Tests

pytest tests/

Project Structure

src/
  └── stinger/
      ├── core/           # Core components and high-level API
      ├── guardrails/     # Guardrail implementations
      ├── data/           # Keyword lists and data files
      ├── scenarios/      # Pre-configured scenarios
      ├── utils/          # Utilities and exceptions
      ├── adapters/       # Model adapters
      ├── cli.py          # CLI entry point
      └── ...

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🆘 Support

• 📖 Documentation
• 🐛 Issues

---

Made with ❤️ for safer AI applications