Overview
Guardrails is a Python framework that helps build reliable AI applications by detecting, quantifying, and mitigating specific types of risks through Input/Output Guards. It provides automatic validation of LLM outputs with pre-built measures called validators.
Key Features
- Guardrails Hub: Collection of pre-built validators for specific risk types
- Input/Output Validation: Every prompt and response passes through validation layers
- Streaming Support: Validates fragments in real-time as LLM returns them
- Composable Guards: Combine multiple validators into Input and Output Guards
Common Validators
- PII Detection: Detect personally identifiable information in text
- Logical Consistency: Validate logical consistency and detect/correct fallacies
- Profanity Check: Ensure content is appropriate
- Factuality: Ensure generated content accuracy and reduce brand risk
- Toxicity: Detect harmful or offensive language
How It Works
Guardrails defines an API and specification format for automatic validation:
- Input validation inspects incoming user prompts immediately
- LLM generates response
- Output validation checks response before returning to user
- Invalid outputs can be corrected or rejected based on configuration
Use Cases
- Enterprise AI applications requiring compliance
- Customer-facing chatbots needing content safety
- Document generation with quality requirements
- Applications handling sensitive data
- Brand-safe content generation
Integration
Works with major frameworks including LangChain, LiteLLM, and OpenAI Agents SDK.
Pricing
Free and open-source on GitHub.
