Amazon Bedrock: Generative AI with FsCDK

Amazon Bedrock is a fully managed service that makes foundation models (FMs) from leading AI companies available through a unified API. FsCDK provides type-safe builders for Bedrock Agents, Knowledge Bases, Data Sources, and Guardrails.

What is Amazon Bedrock?

Amazon Bedrock lets you build and scale generative AI applications using foundation models from Anthropic (Claude), Amazon (Titan), Meta (Llama), and others. Key components include:

• Agents - Autonomous AI assistants that can plan and execute tasks
• Knowledge Bases - RAG (Retrieval-Augmented Generation) with your data
• Data Sources - Connect S3, web crawlers, and other data to Knowledge Bases
• Guardrails - Content filtering and safety controls for AI responses

Quick Start

#r "../src/bin/Release/net8.0/publish/Amazon.JSII.Runtime.dll"
#r "../src/bin/Release/net8.0/publish/Constructs.dll"
#r "../src/bin/Release/net8.0/publish/Amazon.CDK.Lib.dll"
#r "../src/bin/Release/net8.0/publish/System.Text.Json.dll"
#r "../src/bin/Release/net8.0/publish/FsCDK.dll"

open Amazon.CDK
open Amazon.CDK.AwsBedrock
open FsCDK

Use Cases

AI Agent with Custom Instructions

stack "AgentStack" {
    description "Bedrock Agent with full configuration"

    let supportAgent =
        bedrockAgent "CustomerSupportAgent" {
            foundationModel "anthropic.claude-3-sonnet-20240229-v1:0"

            instruction
                "You are a customer support agent. Help users with order tracking, returns, and product questions."

            description "Handles customer inquiries"
            agentResourceRoleArn "arn:aws:iam::123456789012:role/BedrockAgentRole"
            idleSessionTtlInSeconds 1800
            autoPrepare true
            tags [ "Team", "AI"; "Environment", "Production" ]
        }

    ()
}

Knowledge Base with RAG

stack "KnowledgeBaseStack" {
    description "Bedrock Knowledge Base for document retrieval"

    let kbConfig =
        CfnKnowledgeBase.KnowledgeBaseConfigurationProperty(
            Type = "VECTOR",
            VectorKnowledgeBaseConfiguration =
                CfnKnowledgeBase.VectorKnowledgeBaseConfigurationProperty(
                    EmbeddingModelArn = "arn:aws:bedrock:us-east-1::foundation-model/amazon.titan-embed-text-v2:0"
                )
        )

    let storageConfig =
        CfnKnowledgeBase.StorageConfigurationProperty(
            Type = "OPENSEARCH_SERVERLESS",
            OpensearchServerlessConfiguration =
                CfnKnowledgeBase.OpenSearchServerlessConfigurationProperty(
                    CollectionArn = "arn:aws:aoss:us-east-1:123456789012:collection/my-collection",
                    FieldMapping =
                        CfnKnowledgeBase.OpenSearchServerlessFieldMappingProperty(
                            MetadataField = "metadata",
                            TextField = "text",
                            VectorField = "vector"
                        ),
                    VectorIndexName = "my-index"
                )
        )

    let productKB =
        bedrockKnowledgeBase "ProductDocsKB" {
            description "Knowledge base for product documentation"
            roleArn "arn:aws:iam::123456789012:role/BedrockKBRole"
            knowledgeBaseConfiguration kbConfig
            storageConfiguration storageConfig
            tags [ "Team", "AI" ]
        }

    ()
}

Data Source from S3

stack "DataSourceStack" {
    description "Bedrock Data Source connected to S3"

    let s3Config =
        CfnDataSource.DataSourceConfigurationProperty(
            Type = "S3",
            S3Configuration = CfnDataSource.S3DataSourceConfigurationProperty(BucketArn = "arn:aws:s3:::my-docs-bucket")
        )

    let docsSource =
        bedrockDataSource "ProductDocsSource" {
            knowledgeBaseId "KBXXXXXXXX"
            description "S3 bucket containing product documentation"
            dataSourceConfiguration s3Config
            dataDeletionPolicy "RETAIN"
        }

    ()
}

Content Safety Guardrail

stack "GuardrailStack" {
    description "Bedrock Guardrail for content safety"

    let safetyGuardrail =
        bedrockGuardrail "ContentSafety" {
            description "Filters harmful and inappropriate content"
            blockedInputMessaging "Your input contains content that is not allowed."
            blockedOutputsMessaging "The response was filtered for safety."

            contentPolicyConfig (
                CfnGuardrail.ContentPolicyConfigProperty(
                    FiltersConfig =
                        [| CfnGuardrail.ContentFilterConfigProperty(
                               Type = "SEXUAL",
                               InputStrength = "HIGH",
                               OutputStrength = "HIGH"
                           )
                           CfnGuardrail.ContentFilterConfigProperty(
                               Type = "VIOLENCE",
                               InputStrength = "HIGH",
                               OutputStrength = "HIGH"
                           )
                           CfnGuardrail.ContentFilterConfigProperty(
                               Type = "HATE",
                               InputStrength = "HIGH",
                               OutputStrength = "HIGH"
                           ) |]
                )
            )

            tags [ "Environment", "Production" ]
        }

    ()
}

Complete Production Example

stack "ProductionBedrockStack" {
    env (
        environment {
            account config.Account
            region config.Region
        }
    )

    description "Production Bedrock AI platform"
    tags [ "Environment", "Production"; "ManagedBy", "FsCDK" ]

    // Content safety guardrail
    let guardrail =
        bedrockGuardrail "ProductionSafety" {
            description "Production content safety filters"
            blockedInputMessaging "This input is not permitted."
            blockedOutputsMessaging "This response has been filtered."

            contentPolicyConfig (
                CfnGuardrail.ContentPolicyConfigProperty(
                    FiltersConfig =
                        [| CfnGuardrail.ContentFilterConfigProperty(
                               Type = "SEXUAL",
                               InputStrength = "HIGH",
                               OutputStrength = "HIGH"
                           )
                           CfnGuardrail.ContentFilterConfigProperty(
                               Type = "VIOLENCE",
                               InputStrength = "HIGH",
                               OutputStrength = "HIGH"
                           ) |]
                )
            )

            topicPolicyConfig (
                CfnGuardrail.TopicPolicyConfigProperty(
                    TopicsConfig =
                        [| CfnGuardrail.TopicConfigProperty(
                               Name = "Financial Advice",
                               Definition = "Providing specific investment or financial planning advice",
                               Type = "DENY"
                           ) |]
                )
            )

            kmsKeyArn "arn:aws:kms:us-east-1:123456789012:key/my-key"
            tags [ "Environment", "Production" ]
        }

    // AI Agent
    let agent =
        bedrockAgent "ProductionAgent" {
            foundationModel "anthropic.claude-3-sonnet-20240229-v1:0"
            instruction "You are a helpful product assistant. Answer questions accurately based on the knowledge base."
            description "Production customer-facing AI agent"
            agentResourceRoleArn "arn:aws:iam::123456789012:role/BedrockAgentRole"
            idleSessionTtlInSeconds 3600
            autoPrepare true
            tags [ "Environment", "Production" ]
        }

    ()
}

Best Practices

Security

• Use KMS encryption for guardrails and agents handling sensitive data
• Apply least-privilege IAM roles for agent resource roles
• Enable guardrails on all customer-facing agents to prevent harmful content
• Use topic policies to prevent the agent from discussing prohibited subjects

Cost

• Choose the right foundation model for your use case (smaller models for simpler tasks)
• Set appropriate idle session TTL to avoid unnecessary session costs
• Use data deletion policies (RETAIN vs DELETE) based on compliance needs

Reliability

• Set autoPrepare to true (default) so agents are ready after deployment
• Use content filters at HIGH strength for production workloads
• Test guardrails thoroughly before deploying to production

Operational Excellence

• Tag all Bedrock resources for cost allocation and tracking
• Use separate agents for different use cases rather than one monolithic agent
• Version your knowledge base data sources for reproducibility

Default Settings

Escape Hatch

Access the underlying CDK resources via the mutable Agent, KnowledgeBase, DataSource, or Guardrail properties on the spec records after stack synthesis for advanced scenarios not covered by these builders.

Learning Resources

• Amazon Bedrock Developer Guide - Official AWS documentation
• Bedrock Agents - Building autonomous AI agents
• Knowledge Bases for Bedrock - RAG with your data
• Bedrock Guardrails - Content safety controls