Agent Guardrails

LLM-based output validation and quality assurance for task outputs

Overview

Guardrails provide a powerful validation and quality assurance layer for task outputs in PraisonAI. They allow you to define validation criteria that are checked against task results, ensuring outputs meet specific requirements before being accepted.

Diagram

Types of Guardrails

Function Guardrails

Python functions that programmatically validate outputs with precise control

LLM Guardrails

Use natural language criteria evaluated by an LLM for flexible validation

Quick Start

Function Guardrail

python

from praisonaiagents import Agent, Task

def validate_length(output):
    """Ensure output is between 100-500 characters"""
    length = len(output.raw)
    if 100 <= length <= 500:
        return (True, output)  # Pass: (success, result)
    else:
        return (False, f"Output must be 100-500 chars, got {length}")  # Fail: (success, error)

task = Task(
    description="Write a product description",
    agent=agent,
    guardrails=validate_length
)

String Guardrail

python

from praisonaiagents import Agent, Task

# LLM evaluates this criteria automatically
task = Task(
    description="Write marketing copy",
    agent=agent,
    guardrails="Check if professional, no offensive language, includes 3+ benefits"
)

API Reference

GuardrailResult

The return type for guardrail validation functions.

python

class GuardrailResult:
    success: bool   # Whether the validation passed
    result: Any     # Modified output if any, or None
    error: str      # Error message if failed (default: "")

LLMGuardrail

Creates an LLM-based guardrail for natural language validation.

python

LLMGuardrail(
    criteria: str,                    # Natural language validation criteria
    llm_model: str = "gpt-4",        # LLM model to use
    temperature: float = 0.1,         # Low temperature for consistency
    max_tokens: int = 100            # Token limit for response
)

Task Parameters

Configure guardrails on tasks:

python

Task(
    description: str,
    agent: Agent,
    guardrails: Union[Callable, LLMGuardrail, str, None] = None,
    max_retries: int = 3,  # Max retries if guardrail fails
    expected_output: str = None
)

Implementation Examples

Basic Validation

python

def validate_email_format(output: str) -> GuardrailResult:
    """Ensure output contains valid email addresses"""
    import re
    email_pattern = r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b'

    if re.search(email_pattern, output):
        return GuardrailResult(
            success=True,
            error="Valid email found in output"
        )
    else:
        return GuardrailResult(
            success=False,
            error="Output must contain at least one valid email address"
        )

# Use with task
task = Task(
    description="Extract contact emails from the document",
    agent=data_agent,
    guardrails=validate_email_format
)

Complex Structure Validation

python

def validate_json_structure(output: str) -> GuardrailResult:
    """Validate JSON output has required fields"""
    import json

    try:
        data = json.loads(output)
        required_fields = ["title", "summary", "tags", "date"]

        missing = [f for f in required_fields if f not in data]
        if missing:
            return GuardrailResult(
                success=False,
                message=f"Missing required fields: {', '.join(missing)}"
            )

        if not isinstance(data.get("tags"), list):
            return GuardrailResult(
                success=False,
                error="'tags' must be a list"
            )

        return GuardrailResult(
            success=True,
            error="JSON structure is valid"
        )

    except json.JSONDecodeError as e:
        return GuardrailResult(
            success=False,
            message=f"Invalid JSON: {str(e)}"
        )

LLM-based Content Validation

python

# Create sophisticated content guardrail
content_guardrail = LLMGuardrail(
    criteria="""
    Evaluate the output based on:

    1. Accuracy: Information should be factually correct
    2. Completeness: All requested points should be addressed
    3. Clarity: Language should be clear and unambiguous
    4. Tone: Professional but approachable
    5. Structure: Well-organized with clear sections

    The output should NOT contain:
    - Personal opinions presented as facts
    - Promotional language or bias
    - Technical jargon without explanation
    - Grammatical or spelling errors

    Rate as VALID only if all criteria are met.
    """,
    llm_model="gpt-4",
    temperature=0.1
)

# Apply to content generation task
task = Task(
    description="Write a technical blog post about cloud computing",
    agent=writer_agent,
    guardrails=content_guardrail,
    max_retries=2
)

Output Modification Guardrails

python

def sanitize_pii(output: str) -> GuardrailResult:
    """Remove personally identifiable information"""
    import re

    # Check for PII patterns
    ssn_pattern = r'\b\d{3}-\d{2}-\d{4}\b'
    if re.search(ssn_pattern, output):
        # Modify output to remove PII
        sanitized = re.sub(ssn_pattern, '[SSN-REDACTED]', output)
        return GuardrailResult(
            success=True,
            message=f"PII removed. Modified output: {sanitized}"
        )

    return GuardrailResult(
        success=True,
        error="No PII detected"
    )

Creating an Implementation Guide

Step 1: Identify Validation Needs

Determine what aspects of the output need validation:

Format requirements (JSON, XML, Markdown)
Content requirements (specific information, tone)
Security concerns (no sensitive data, safe content)
Business rules (pricing limits, compliance)

Step 2: Choose Guardrail Type

Use Function Guardrails When

Use LLM Guardrails When

Validation rules are clear and programmatic

Validation requires judgment or context

Performance is critical

Natural language criteria are easier to define

Deterministic results needed

Flexibility in interpretation is valuable

Checking formats, patterns, or calculations

Evaluating quality, tone, or completeness

Step 3: Implement Guardrails

python

from praisonaiagents import Agent, Task, GuardrailResult, LLMGuardrail

# Example: E-commerce product description validation
def validate_product_description(output: str) -> GuardrailResult:
    """Validate product descriptions meet requirements"""

    # Check minimum length
    if len(output) < 50:
        return GuardrailResult(
            success=False,
            error="Description too short (min 50 characters)"
        )

    # Check for required elements
    required_words = ["features", "benefits", "specifications"]
    missing = [w for w in required_words if w.lower() not in output.lower()]

    if missing:
        return GuardrailResult(
            success=False,
            message=f"Missing sections: {', '.join(missing)}"
        )

    # Check for price format
    import re
    price_pattern = r'\$\d+\.?\d{0,2}'
    if not re.search(price_pattern, output):
        return GuardrailResult(
            success=False,
            error="Must include price in format $XX.XX"
        )

    return GuardrailResult(
        success=True,
        error="Product description meets all requirements"
    )

# Combine with LLM guardrail for quality
quality_guardrail = LLMGuardrail(
    criteria="""
    Check if the product description:
    1. Is engaging and persuasive
    2. Highlights unique selling points
    3. Uses active voice
    4. Avoids superlatives without evidence
    5. Includes sensory details where appropriate
    """
)

# Create agent and tasks
product_agent = Agent(
    name="ProductWriter",
    role="E-commerce Copywriter"
)

# Task with multiple validation steps
description_task = Task(
    description="Write product description for wireless headphones",
    agent=product_agent,
    guardrails=validate_product_description  # Format validation
)

quality_task = Task(
    description="Enhance the description for engagement",
    agent=product_agent,
    guardrails=quality_guardrail,  # Quality validation
    context=[description_task]
)

Best Practices

Clear Validation Criteria

Be specific about what constitutes valid output
Provide examples of valid and invalid outputs
Consider edge cases and exceptions
Document validation rules clearly

Balanced Strictness

Too strict: May cause excessive retries
Too lenient: May accept poor quality outputs
Test with various inputs to find the right balance
Consider allowing partial success where appropriate

Helpful Error Messages

Clearly explain why validation failed
Suggest how to fix the issue
Include specific examples when helpful
Avoid generic error messages

Retry Configuration

Default is 3 retries, adjust based on task complexity
Consider the cost of retries (API calls, time)
Some tasks may benefit from no retries
Log retry attempts for debugging

Common Use Cases

Content Moderation

python

moderation_guardrail = LLMGuardrail(
    criteria="""
    Ensure the content is appropriate for all audiences:
    - No profanity or offensive language
    - No discriminatory or biased statements  
    - No violence or graphic descriptions
    - Maintains respectful tone throughout
    """
)

Data Format Validation

python

def validate_csv_output(output: str) -> GuardrailResult:
    """Ensure output is valid CSV with required columns"""
    import csv
    import io

    try:
        reader = csv.DictReader(io.StringIO(output))
        required_columns = ["name", "email", "phone", "address"]

        if not all(col in reader.fieldnames for col in required_columns):
            return GuardrailResult(
                success=False,
                message=f"Missing columns. Required: {required_columns}"
            )

        return GuardrailResult(success=True, error="Valid CSV format")

    except Exception as e:
        return GuardrailResult(
            success=False,
            message=f"Invalid CSV: {str(e)}"
        )

Compliance Checking

python

compliance_guardrail = LLMGuardrail(
    criteria="""
    Verify the output complies with GDPR requirements:
    1. Includes privacy policy reference
    2. Mentions data retention period
    3. Explains user rights (access, deletion, portability)
    4. Identifies data controller
    5. No unnecessary personal data collection
    """
)

Quality Control

python

def validate_code_quality(output: str) -> GuardrailResult:
    """Check generated code meets quality standards"""

    # Check for common issues
    if "eval(" in output or "exec(" in output:
        return GuardrailResult(
            success=False,
            error="Unsafe functions (eval/exec) detected"
        )

    # Check indentation
    lines = output.split('\n')
    if any(line.startswith(' ') and len(line) - len(line.lstrip()) % 4 != 0 
           for line in lines):
        return GuardrailResult(
            success=False,
            error="Inconsistent indentation (use 4 spaces)"
        )

    return GuardrailResult(
        success=True,
        error="Code meets quality standards"
    )

Real-world Example

Financial Report Validation

python

from praisonaiagents import Agent, Task, AgentTeam, GuardrailResult, LLMGuardrail
import json
import re

def validate_financial_report(output: str) -> GuardrailResult:
    """Validate financial report structure and calculations"""
    try:
        report = json.loads(output)

        # Check required sections
        required_sections = ["revenue", "expenses", "profit", "summary"]
        missing = [s for s in required_sections if s not in report]
        if missing:
            return GuardrailResult(
                success=False,
                message=f"Missing sections: {missing}"
            )

        # Validate calculations
        revenue = report.get("revenue", {}).get("total", 0)
        expenses = report.get("expenses", {}).get("total", 0)
        stated_profit = report.get("profit", {}).get("total", 0)
        calculated_profit = revenue - expenses

        if abs(stated_profit - calculated_profit) > 0.01:
            return GuardrailResult(
                success=False,
                message=f"Profit calculation error: stated {stated_profit}, should be {calculated_profit}"
            )

        # Check for negative values where inappropriate
        if revenue < 0:
            return GuardrailResult(
                success=False,
                error="Revenue cannot be negative"
            )

        return GuardrailResult(
            success=True,
            error="Financial report validated successfully"
        )

    except Exception as e:
        return GuardrailResult(
            success=False,
            message=f"Report validation error: {str(e)}"
        )

# LLM guardrail for report quality
report_quality_guardrail = LLMGuardrail(
    criteria="""
    Evaluate the financial report for:
    1. Clear executive summary with key insights
    2. Proper context for all numbers (YoY growth, percentages)
    3. Risk factors and mitigation strategies mentioned
    4. Future outlook based on current trends
    5. Professional language appropriate for stakeholders
    6. No speculation presented as fact
    """
)

# Create workflow
analyst = Agent(
    name="FinancialAnalyst",
    role="Senior Financial Analyst",
    goal="Create accurate financial reports"
)

# Multi-step validation
generate_report = Task(
    description="Generate Q4 financial report from data",
    agent=analyst,
    guardrails=validate_financial_report,  # Structural validation
    max_retries=2
)

enhance_report = Task(
    description="Add analysis and insights to the report",
    agent=analyst,
    guardrails=report_quality_guardrail,  # Quality validation
    context=[generate_report]
)

# Execute workflow
workflow = AgentTeam(
    agents=[analyst],
    tasks=[generate_report, enhance_report]
)

result = workflow.start()

Summary

Guardrails in PraisonAI provide:

✅ Output Validation – Ensure task outputs meet specific criteria ✅ Quality Assurance – Maintain consistent quality standards ✅ Error Prevention – Catch issues before they propagate ✅ Flexible Implementation – Use functions or LLMs for validation ✅ Automatic Retries – Built-in retry mechanism for failed validations

Use guardrails whenever you need:

Structured output validation
Content quality checks
Compliance verification
Security and safety checks
Business rule enforcement