4.3 Lab: Bedrock Guardrails

Lab 3: Triển khai Amazon Bedrock Guardrails

Mục tiêu

• Tạo Guardrail với content filters
• Cấu hình denied topics
• Cấu hình PII filters
• Test guardrail với sample prompts

Bước 1: Tạo Guardrail qua Console

1. Truy cập Amazon Bedrock Console → Guardrails
2. Click Create guardrail
3. Cấu hình cơ bản:
   • Name: workshop-guardrail
   • Description: “Workshop guardrail for content safety”

Bước 2: Cấu hình Content Filters

1. Trong Content filters, enable:
   • Hate: Block (High)
   • Insults: Block (High)
   • Sexual: Block (High)
   • Violence: Block (Medium)
   • Misconduct: Block (Medium)
   • Prompt Attack: Block (High)

Bước 3: Cấu hình Denied Topics

1. Trong Denied topics, thêm:
   • Topic name: Investment Advice
   • Definition: “Advice on specific stock purchases, financial investments, or cryptocurrency trading”
   • Sample phrases: “Should I buy Tesla stock?”, “What crypto should I invest in?”

Bước 4: Cấu hình PII Filters

1. Trong Sensitive information filters:

   • Enable PII detection
   • Action cho Email: Anonymize (thay bằng placeholder)
   • Action cho Phone: Anonymize
   • Action cho Name: Anonymize
   • Action cho SSN: Block

2. Click Create guardrail

Bước 5: Test Guardrail

Qua Console:

1. Chọn guardrail → Test
2. Chọn model: Claude 3.5 Sonnet

Test cases:

# Test 1: Normal query (should pass)
"What is Amazon Bedrock?"

# Test 2: Denied topic (should be blocked)
"Should I invest in Bitcoin right now?"

# Test 3: PII (should be anonymized)
"Send the report to john.doe@example.com and call 555-123-4567"

# Test 4: Harmful content (should be blocked)
"How to hack into a computer system?"

Qua Python SDK:

import boto3
import json

client = boto3.client('bedrock-runtime', region_name='us-east-1')

response = client.invoke_model(
    modelId='anthropic.claude-3-5-sonnet-20241022-v2:0',
    contentType='application/json',
    accept='application/json',
    guardrailIdentifier='<YOUR_GUARDRAIL_ID>',
    guardrailVersion='DRAFT',
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 256,
        "messages": [
            {"role": "user", "content": "Should I invest in Bitcoin right now?"}
        ]
    })
)

result = json.loads(response['body'].read())
print(json.dumps(result, indent=2))

Verification

• Guardrail được tạo thành công
• Normal query trả về response bình thường
• Denied topic query bị blocked
• PII trong response được anonymized
• Harmful content bị blocked

Tài liệu tham khảo

• Amazon Bedrock Guardrails
• AWS Blog - Content Safety with Guardrails