🎯 What are Guardrails?

Guardrails are safety validators that analyze AI responses before they reach users. They calculate metrics (like pii_risk or compliance_score) and block responses that violate your policies.

🏗️ Architecture: 3-Tier System

🤖 LLM-Assisted vs Pattern-Based

🛡️ Security: Defense-in-Depth

1. Prompt Injection Detection - Blocks adversarial attacks before LLM call
2. LLM Response Validation - Verifies LLM output integrity
3. Pattern Fallback - Always available if LLM fails
4. Input Sanitization - Removes malicious control characters

💡 Best Practices

• ✅ Use pattern-based guardrails first - PII, profanity → regex/keyword matching
• ⚠️ LLM for complex cases only - Financial/medical compliance
• 🔄 Always provide fallback keywords - System works without OPENAI_API_KEY
• 🎯 Set appropriate thresholds - 90+ strict, 70-80 balanced, <70 lenient
• 🧪 Test before deploying - Use the Tester tab with real scenarios

🚀 Deployment Workflow

📚 API Reference

For programmatic access, see the API Documentation.

POST /api/guardrails/register
{
  "id": "custom_policy_v1",
  "name": "Custom Policy",
  "description": "...",
  "prompt_template": "...",
  "keywords": ["..."],
  "metric_name": "compliance_score",
  "threshold": 75
}