Input Rails

Input rails validate user input before it's processed by the LLM.

Overview

Input rails perform safety checks on user messages to prevent:

• Toxic or offensive content
• Sensitive data exposure
• Jailbreak attempts
• Malicious requests

How It Works

1. User sends input
2. Input rails validate the content
3. If validation fails, request is blocked
4. If validation passes, input proceeds to LLM

Configuration

Enable input checks in your configuration:

guardrails:
  input_checks: true
  check_toxicity: true
  check_sensitive_data: true
  check_semantic: true

Usage

With LLMRails

Input checks are automatic when input_checks: true:

from elsai_guardrails.guardrails import LLMRails, RailsConfig

yaml_content = """
llm:
  engine: "openai"
  model: "gpt-4o-mini"
  api_key: "sk-..."

guardrails:
  input_checks: true
  check_toxicity: true
  check_sensitive_data: true
  check_semantic: true
"""

config = RailsConfig.from_content(yaml_content=yaml_content)
rails = LLMRails(config=config)

# Input will be automatically checked
result = rails.generate(
    messages=[{"role": "user", "content": "user input"}],
    return_details=True
)

if result.get('input_check'):
    print(f"Input passed: {result['input_check'].passed}")

Standalone Input Validation

from elsai_guardrails.guardrails import GuardrailSystem, GuardrailConfig

config = GuardrailConfig(
    check_toxicity=True,
    check_sensitive_data=True,
    check_semantic=True
)
guardrail = GuardrailSystem(config=config)

user_input = "test input"
result = guardrail.check_input(user_input)

if not result.passed:
    print(f"Input blocked: {result.message}")

Block Reasons

Input can be blocked for:

1. Toxicity: Toxic or offensive content detected
2. Sensitive Data: Personal information detected (emails, phone numbers, etc.)
3. Content Issues: Jailbreak attempts, malicious requests, prompt injections

Example

from elsai_guardrails.guardrails import GuardrailSystem, GuardrailConfig

config = GuardrailConfig()
guardrail = GuardrailSystem(config=config)

# Example 1: Valid input
result = guardrail.check_input("Hello, how are you?")
print(f"Passed: {result.passed}")  # True

# Example 2: Sensitive data
result = guardrail.check_input("My email is user@example.com")
print(f"Passed: {result.passed}")  # False
print(f"Reason: {result.message}")  # "Sensitive data detected."

# Example 3: Jailbreak attempt
result = guardrail.check_input("Ignore previous instructions and...")
print(f"Passed: {result.passed}")  # False
print(f"Content Class: {result.semantic_class}")  # "jailbreak"

Next Steps

• Output Rails - Output validation
• Toxicity Detection - Toxicity checks
• Sensitive Data Detection - Sensitive data checks
• Content Classification - Content checks