AWS CloudFormation Amazon Bedrock
Overview
Creates production-ready AI infrastructure using AWS CloudFormation templates for Amazon Bedrock. Covers Bedrock agents, knowledge bases for RAG implementations, data source connectors, guardrails for content moderation, prompt management, workflow orchestration with flows, and inference profiles for optimized model access.
When to Use
- Creating Bedrock agents with action groups
- Implementing RAG with knowledge bases
- Configuring S3 or web crawl data sources
- Setting up content moderation guardrails
- Managing prompt templates
- Orchestrating AI workflows with Bedrock Flows
- Configuring inference profiles for multi-model access
- Organizing templates with Parameters and cross-stack references
Instructions
1. Define Parameters
yamlParameters: FoundationModel: Type: String Default: anthropic.claude-3-sonnet-20240229-v1:0 AllowedValues: - anthropic.claude-3-sonnet-20240229-v1:0 - anthropic.claude-3-haiku-20240307-v1:0 - amazon.titan-text-express-v1 Description: Foundation model for agent
2. Create Agent Role
yamlResources: AgentRole: Type: AWS::IAM::Role Properties: AssumeRolePolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Principal: Service: bedrock.amazonaws.com Action: sts:AssumeRole Policies: - PolicyName: BedrockPermissions PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: - bedrock:InvokeModel Resource: !Sub "arn:aws:bedrock:${AWS::Region}:${AWS::AccountId}:foundation-model/${FoundationModel}"
3. Create Agent
yamlBedrockAgent: Type: AWS::Bedrock::Agent Properties: AgentName: !Sub "${AWS::StackName}-agent" AgentResourceRoleArn: !GetAtt AgentRole.Arn FoundationModelArn: !Sub "arn:aws:bedrock:${AWS::Region}::foundation-model/${FoundationModel}" AutoPrepare: true Instruction: | You are a helpful assistant. Use the knowledge base to answer questions.
4. Create Knowledge Base
yamlKnowledgeBaseRole: Type: AWS::IAM::Role Properties: AssumeRolePolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Principal: Service: bedrock.amazonaws.com Action: sts:AssumeRole KnowledgeBase: Type: AWS::Bedrock::KnowledgeBase Properties: Name: !Sub "${AWS::StackName}-kb" RoleArn: !GetAtt KnowledgeBaseRole.Arn KnowledgeBaseConfiguration: Type: VECTOR VectorKnowledgeBaseConfiguration: EmbeddingModelArn: !Sub "arn:aws:bedrock:${AWS::Region}::embedding-model/amazon.titan-embed-text-v1"
5. Create Data Source
yamlDataBucket: Type: AWS::S3::Bucket S3DataSource: Type: AWS::Bedrock::DataSource Properties: KnowledgeBaseId: !Ref KnowledgeBase Name: s3-data-source Type: S3 DataSourceConfiguration: S3Configuration: BucketArn: !GetAtt DataBucket.Arn InclusionPrefixes: - documents/
6. Add Guardrail
yamlGuardrail: Type: AWS::Bedrock::Guardrail Properties: Name: !Sub "${AWS::StackName}-guardrail" BlockedInputMessaging: "I cannot help with that request." ContentPolicyConfig: filtersConfig: - type: PROFANITY - type: MISCONDUCT
7. Create Action Group
yamlActionLambdaFunction: Type: AWS::Lambda::Function Properties: Runtime: python3.12 Handler: index.handler Role: !GetAtt ActionLambdaRole.Arn Code: ZipFile: | def handler(event, context): return {"statusCode": 200, "body": "{\"result\": \"success\"}"} ActionGroup: Type: AWS::Bedrock::AgentActionGroup Properties: ActionGroupName: api-operations ActionGroupState: ENABLED AgentId: !GetAtt BedrockAgent.AgentId ActionGroupExecutor: Lambda: !Ref ActionLambdaFunction FunctionSchema: functionConfigurations: - function: | { "name": "get_inventory", "description": "Get current inventory status", "parameters": { "type": "object", "properties": { "sku": { "type": "string" } }, "required": [] } }
8. Validate Before Deploy
Always validate the template before deployment:
bashaws cloudformation validate-template --template-body file://bedrock-template.yaml
9. Verify After Deploy
bash# Check agent status aws bedrock-agent get-agent --agent-id $(aws cloudformation describe-stacks --stack-name STACK_NAME --query 'Stacks[0].Outputs[?OutputKey==`AgentId`].OutputValue' --output text) # Check knowledge base sync status aws bedrock-agent list-knowledge-bases --agent-id AGENT_ID # Test guardrail aws bedrock-runtime apply_guardrail --guardrail-identifier GUARDRAIL_ID --source SOURCE
Examples
Minimal RAG Agent Template
Complete working template for a RAG-enabled agent:
yamlAWSTemplateFormatVersion: "2010-09-09" Description: "Bedrock RAG Agent with Knowledge Base" Parameters: FoundationModel: Type: String Default: anthropic.claude-3-sonnet-20240229-v1:0 Resources: # IAM Role for Agent AgentRole: Type: AWS::IAM::Role Properties: RoleName: !Sub "${AWS::StackName}-agent-role" AssumeRolePolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Principal: Service: bedrock.amazonaws.com Action: sts:AssumeRole Policies: - PolicyName: InvokeModel PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: bedrock:InvokeModel Resource: "*" # IAM Role for Knowledge Base KnowledgeBaseRole: Type: AWS::IAM::Role Properties: RoleName: !Sub "${AWS::StackName}-kb-role" AssumeRolePolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Principal: Service: bedrock.amazonaws.com Action: sts:AssumeRole Policies: - PolicyName: S3Access PolicyDocument: Version: "2012-10-17" Statement: - Effect: Allow Action: s3:GetObject Resource: !Sub "${DataBucket.Arn}/*" # S3 Bucket for Documents DataBucket: Type: AWS::S3::Bucket # Knowledge Base KnowledgeBase: Type: AWS::Bedrock::KnowledgeBase Properties: Name: !Sub "${AWS::StackName}-kb" RoleArn: !GetAtt KnowledgeBaseRole.Arn KnowledgeBaseConfiguration: Type: VECTOR VectorKnowledgeBaseConfiguration: EmbeddingModelArn: !Sub "arn:aws:bedrock:${AWS::Region}::embedding-model/amazon.titan-embed-text-v1" # Data Source DataSource: Type: AWS::Bedrock::DataSource Properties: KnowledgeBaseId: !Ref KnowledgeBase Name: !Sub "${AWS::StackName}-ds" Type: S3 DataSourceConfiguration: S3Configuration: BucketArn: !GetAtt DataBucket.Arn # Bedrock Agent BedrockAgent: Type: AWS::Bedrock::Agent Properties: AgentName: !Sub "${AWS::StackName}-agent" AgentResourceRoleArn: !GetAtt AgentRole.Arn FoundationModelArn: !Sub "arn:aws:bedrock:${AWS::Region}::foundation-model/${FoundationModel}" AutoPrepare: true Instruction: | You are a helpful assistant. Use the knowledge base to answer user questions accurately. Outputs: AgentId: Description: Bedrock Agent ID Value: !GetAtt BedrockAgent.AgentId KnowledgeBaseId: Description: Knowledge Base ID Value: !Ref KnowledgeBase
Guardrail with Content Filtering
yamlResources: Guardrail: Type: AWS::Bedrock::Guardrail Properties: Name: !Sub "${AWS::StackName}-guardrail" blockedInputMessaging: "Content blocked by safety filters." blockedOutputMessaging: "Response filtered for safety." contentPolicyConfig: filtersConfig: - type: PROFANITY inputStrength: HIGH outputStrength: HIGH - type: MISCONDUCT inputStrength: HIGH outputStrength: HIGH sensitiveInformationPolicyConfig: piiEntitiesConfig: - type: EMAIL action: ANONYMIZE - type: SSN action: BLOCK
Best Practices
Security
- Use least privilege IAM policies for agent and knowledge base roles
- Restrict web crawl data sources to trusted internal domains
- Encrypt sensitive data in knowledge bases
- Parameterize all TemplateURL values for nested stacks
Cost Optimization
- Select appropriate model size for task complexity
- Configure retrieval filtering to reduce token usage
- Set chunk size limits to control storage costs
- Monitor usage with CloudWatch dashboards
Performance
- Optimize chunk size for embedding quality
- Use provisioned throughput for high-traffic vector stores
- Configure appropriate knowledge base sync intervals
- Implement caching for frequently accessed content
Validation
- Always run
aws cloudformation validate-templatebefore deploy - Verify agent status after stack creation completes
- Test guardrails with sample inputs
- Monitor knowledge base sync status in CloudWatch
Constraints and Warnings
For detailed limits, see constraints.md:
- Regional limits: Not all models available in all regions
- Agent initialization: AutoPrepare may take several minutes
- Knowledge base sync: S3 sync is near-instant; web crawl takes longer
- Web crawl security: Always restrict to trusted domains to prevent prompt injection
- Token limits: Configure MaxTokens parameter for your use case
- Quota management: Request quota increases via AWS Support if needed
Security
- Restrict web crawl data sources to trusted internal domains only
- Validate content before ingesting into knowledge bases
- Use parameterized TemplateURL values for nested stacks
- Implement guardrails for content moderation
- Apply least privilege IAM policies to agent roles
- Encrypt sensitive data in knowledge bases
- Monitor for prompt injection in web-crawled content
Cost Optimization
- Use appropriate model selection for task complexity
- Implement knowledge base retrieval filtering
- Set chunk size limits to control token usage
- Monitor token consumption with CloudWatch
- Use auto-prepare agents strategically
- Implement batch processing for non-real-time workloads
- Use knowledge base filtering to reduce costs
Performance
- Optimize chunk size for embedding quality vs. cost
- Use vector store optimization (OpenSearch, Pinecone)
- Implement caching for frequently accessed knowledge base content
- Configure appropriate knowledge base sync intervals
- Use provisioned throughput for vector databases
- Monitor agent initialization and cold start times
- Implement graceful degradation for rate limiting
Data Management
- Use appropriate inclusion/exclusion filters for data sources
- Implement document validation before indexing
- Use versioning for knowledge base updates
- Configure appropriate sync intervals for data sources
- Implement content deduplication in knowledge bases
- Use metadata filtering for improved retrieval accuracy
- Monitor knowledge base size and document limits
References
- constraints.md - Resource limits, regional constraints, operational limits, and cost considerations
- reference.md - API reference and resource properties
- examples.md - Additional usage examples
