GuardrailResult

The GuardrailResult dataclass contains the results of guardrail checks.

Overview

GuardrailResult is returned by all guardrail check methods and provides comprehensive information about the validation results.

Attributes

passed

Type: bool

Whether all guardrail checks passed.

if result.passed:
    print("Content is safe")
else:
    print("Content was blocked")

toxicity

Type: Dict[str, Any]

Toxicity detection results. Contains:

• label: Predicted class ('toxic', 'offensive', 'non-toxic')
• confidence: Confidence score (0.0-1.0)

toxicity = result.toxicity
print(f"Label: {toxicity.get('label')}")
print(f"Confidence: {toxicity.get('confidence')}")

sensitive_data

Type: Dict[str, Any]

Sensitive data detection results. Contains:

• predicted_labels: List of detected sensitive data types

sensitive = result.sensitive_data
labels = sensitive.get('predicted_labels', [])
if labels and labels != ["No sensitive data detected"]:
    print(f"Detected: {labels}")

semantic_class

Type: str

Content classification result. Possible values:

• "jailbreak": Jailbreak attempt detected
• "malicious": Malicious content detected
• "prompt_injection": Prompt injection detected
• "malicious_code_injection": Code injection detected
• "misc": Other problematic content
• Empty string if no semantic issues detected

if result.semantic_class:
    print(f"Content class: {result.semantic_class}")

message

Type: str

Human-readable message describing the result.

print(result.message)

Example Usage

Basic Check

from elsai_guardrails.guardrails import GuardrailSystem, GuardrailConfig

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

result = guardrail.check_text("Hello, this is a test")

print(f"Passed: {result.passed}")
print(f"Message: {result.message}")

Detailed Analysis

result = guardrail.check_input("test input")

# Check overall result
if not result.passed:
    print(f"Blocked: {result.message}")
    
    # Check toxicity
    if result.toxicity:
        toxicity_label = result.toxicity.get('label')
        toxicity_conf = result.toxicity.get('confidence')
        if toxicity_label in ['toxic', 'offensive']:
            print(f"Toxicity: {toxicity_label} (confidence: {toxicity_conf})")
    
    # Check sensitive data
    if result.sensitive_data:
        labels = result.sensitive_data.get('predicted_labels', [])
        if labels and labels != ["No sensitive data detected"]:
            print(f"Sensitive data: {labels}")
    
    # Check content class
    if result.semantic_class:
        print(f"Content issue: {result.semantic_class}")

From LLMRails

from elsai_guardrails.guardrails import LLMRails, RailsConfig

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

result = rails.generate(
    messages=[{"role": "user", "content": "test"}],
    return_details=True
)

# Access input check result
if result.get('input_check'):
    input_result = result['input_check']
    print(f"Input passed: {input_result.passed}")
    print(f"Toxicity: {input_result.toxicity}")
    print(f"Sensitive: {input_result.sensitive_data}")
    print(f"Semantic: {input_result.semantic_class}")

# Access output check result
if result.get('output_check'):
    output_result = result['output_check']
    print(f"Output passed: {output_result.passed}")

Result Examples

Passed Result

GuardrailResult(
    passed=True,
    toxicity={'label': 'non-toxic', 'confidence': 0.1},
    sensitive_data={'predicted_labels': ['No sensitive data detected']},
    semantic_class='',
    message='All checks passed'
)

Blocked Result (Toxicity)

GuardrailResult(
    passed=False,
    toxicity={'label': 'toxic', 'confidence': 0.85},
    sensitive_data={'predicted_labels': ['No sensitive data detected']},
    semantic_class='',
    message='Toxic content detected.'
)

Blocked Result (Sensitive Data)

GuardrailResult(
    passed=False,
    toxicity={'label': 'non-toxic', 'confidence': 0.1},
    sensitive_data={'predicted_labels': ['Email']},
    semantic_class='',
    message='Sensitive data detected.'
)

Blocked Result (Content Classification)

GuardrailResult(
    passed=False,
    toxicity={'label': 'non-toxic', 'confidence': 0.1},
    sensitive_data={'predicted_labels': ['No sensitive data detected']},
    semantic_class='jailbreak',
    message='Jailbreak attempt detected.'
)

Next Steps

• GuardrailSystem - Using guardrail checks
• LLMRails - High-level API
• Python API - More examples