Bedrock Guardrails: Preventing AI Hallucinations in Production

Bedrock Guardrails: Preventing AI Hallucinations in Production

AI hallucinations cost companies money and trust. Amazon Bedrock Guardrails provides enterprise-grade controls to filter harmful content, enforce topic restrictions, and ground responses in facts. Here’s how to implement them.

Guardrails Capabilities

Creating a Guardrail

import boto3

bedrock = boto3.client('bedrock')

# Create guardrail
guardrail = bedrock.create_guardrail(
    name='customer-support-guardrail',
    description='Guardrails for customer-facing chatbot',

    # Content filtering
    contentPolicyConfig={
        'filtersConfig': [
            {'type': 'HATE', 'inputStrength': 'HIGH', 'outputStrength': 'HIGH'},
            {'type': 'VIOLENCE', 'inputStrength': 'HIGH', 'outputStrength': 'HIGH'},
            {'type': 'SEXUAL', 'inputStrength': 'HIGH', 'outputStrength': 'HIGH'},
        ]
    },

    # Topic restrictions
    topicPolicyConfig={
        'topicsConfig': [
            {
                'name': 'Competitor Discussion',
                'definition': 'Any mention of competitor products or pricing',
                'type': 'DENY'
            },
            {
                'name': 'Financial Advice',
                'definition': 'Investment or financial planning recommendations',
                'type': 'DENY'
            }
        ]
    },

    # PII handling
    sensitiveInformationPolicyConfig={
        'piiEntitiesConfig': [
            {'type': 'US_SOCIAL_SECURITY_NUMBER', 'action': 'ANONYMIZE'},
            {'type': 'CREDIT_DEBIT_CARD_NUMBER', 'action': 'BLOCK'},
        ]
    },

    blockedInputMessaging='I cannot help with that request.',
    blockedOutputsMessaging='I cannot provide that information.'
)

Using Guardrails with Invoke

runtime = boto3.client('bedrock-runtime')

response = runtime.invoke_model(
    modelId='amazon.nova-pro-v1:0',
    guardrailIdentifier='my-guardrail-id',
    guardrailVersion='DRAFT',
    body=json.dumps({
        'inputText': user_question
    })
)

# Check if guardrail intervened
if response.get('guardrailAction') == 'INTERVENED':
    print("Guardrail blocked this request")

Grounding Checks (Anti-Hallucination)

# Enable grounding to verify facts
response = runtime.apply_guardrail(
    guardrailIdentifier='my-guardrail',
    guardrailVersion='1',
    source='OUTPUT',
    content=[{
        'text': {
            'text': model_response,
            'qualifiers': ['grounding_source']
        }
    }],
    # Provide source documents for verification
    groundingSource={
        'text': source_documents
    }
)