Guardrails

Guardrails are a safety and validation system in langcrew that ensures data quality, security, and compliance. They act as protective barriers that validate inputs before processing and outputs before delivery.

What are Guardrails?

Guardrails are validation functions that can be applied to:

• Inputs: Validate incoming data before processing
• Outputs: Validate generated content before delivery
• Agents: Apply global validation rules to all agent operations
• Tasks: Apply specific validation rules to individual tasks

Think of guardrails as quality control checkpoints that ensure your AI system operates safely and within defined boundaries.

Core Concepts

Input vs Output Guardrails

Input Guardrails validate data before it reaches your AI agents:

• Detect sensitive information (passwords, API keys, PII)
• Validate data formats and structures
• Check user permissions and access rights
• Implement rate limiting and abuse prevention

Output Guardrails validate content after AI processing:

• Ensure content meets quality standards
• Filter inappropriate or harmful content
• Validate output format and structure
• Check for factual accuracy and balanced language

Multi-Layer Protection

Guardrails can be applied at multiple levels:

• Agent-Level: Apply to all tasks executed by an agent
• Task-Level: Apply only to specific tasks
• Combined: Layer multiple guardrail types for comprehensive protection

Basic Structure

Every guardrail function follows a simple pattern:

@input_guard  # or @output_guard
def my_guardrail(data: Any) -> Tuple[bool, str]:
    """
    Args:
        data: The data to validate

    Returns:
        Tuple[bool, str]: (is_valid, message)
    """
    # Validation logic here
    if validation_passes:
        return True, "✅ Validation passed"
    else:
        return False, "❌ Validation failed: reason"

Key Components

Decorators

• @input_guard: Marks functions as input validators
• @output_guard: Marks functions as output validators

Error Handling

When guardrails fail, they raise GuardrailError exceptions with detailed information about what went wrong.

Integration Points

Guardrails integrate seamlessly with:

• Agents: Apply via input_guards and output_guards parameters
• Tasks: Apply specific guardrails to individual tasks
• Crews: Orchestrate tasks and agents with their respective guardrails

Benefits

• Safety: Prevent processing of sensitive or harmful data
• Quality: Ensure outputs meet your standards
• Compliance: Enforce business rules and regulations
• Reliability: Catch errors early in the processing pipeline
• Flexibility: Mix and match guardrails for different use cases

Use Cases

• Content Generation: Validate AI-generated content quality
• Data Processing: Ensure input data meets requirements
• Security: Prevent exposure of sensitive information
• Compliance: Enforce industry-specific regulations
• Quality Control: Maintain consistent output standards