Skip to content

Latest commit

 

History

History
1470 lines (1169 loc) · 50.5 KB

README.md

File metadata and controls

1470 lines (1169 loc) · 50.5 KB
Raw

Amazon Bedrock Construct Library

Stability: Experimental

All classes are under active development and subject to non-backward compatible changes or removal in any future version. These are not subject to the Semantic Versioning model. This means that while you may use them, you may need to update your source code when upgrading to a newer version of this package.

Language Package
Typescript Logo TypeScript @cdklabs/generative-ai-cdk-constructs
Python Logo Python cdklabs.generative_ai_cdk_constructs

Amazon Bedrock is a fully managed service that offers a choice of foundation models (FMs) along with a broad set of capabilities for building generative AI applications.

This construct library includes CloudFormation L1 resources to deploy Bedrock features.

Table of contents

API

See the API documentation.

Knowledge Bases

With Knowledge Bases for Amazon Bedrock, you can give FMs and agents contextual information from your company’s private data sources for Retrieval Augmented Generation (RAG) to deliver more relevant, accurate, and customized responses.

Create a Knowledge Base

A vector index on a vector store is required to create a Knowledge Base. This construct currently supports Amazon OpenSearch Serverless, Amazon RDS Aurora PostgreSQL, Pinecone . By default, this resource will create an OpenSearch Serverless vector collection and index for each Knowledge Base you create, but you can provide an existing collection and/or index to have more control. For other resources you need to have the vector stores already created and credentials stored in AWS Secrets Manager. For Aurora, the construct provides an option to create a default AmazonAuroraDefaultVectorStore construct that will provision the vector store backed by Amazon Aurora for you. To learn more you can read here.

The resource accepts an instruction prop that is provided to any Bedrock Agent it is associated with so the agent can decide when to query the Knowledge Base.

Amazon Bedrock Knowledge Bases currently only supports S3 as a data source. The S3DataSource resource is used to configure how the Knowledge Base handles the data source.

Example of OpenSearch Serverless:

TypeScript

import * as s3 from "aws-cdk-lib/aws-s3";
import { bedrock } from "@cdklabs/generative-ai-cdk-constructs";

const kb = new bedrock.KnowledgeBase(this, "KnowledgeBase", {
  embeddingsModel: bedrock.BedrockFoundationModel.TITAN_EMBED_TEXT_V1,
  instruction: "Use this knowledge base to answer questions about books. " + "It contains the full text of novels.",
});

const docBucket = new s3.Bucket(this, "DocBucket");

new bedrock.S3DataSource(this, "DataSource", {
  bucket: docBucket,
  knowledgeBase: kb,
  dataSourceName: "books",
  chunkingStrategy: bedrock.ChunkingStrategy.fixedSize({
    maxTokens: 500,
    overlapPercentage: 20,
  }),
});

Python

from aws_cdk import (
    aws_s3 as s3,
)
from cdklabs.generative_ai_cdk_constructs import (
    bedrock
)

kb = bedrock.KnowledgeBase(self, 'KnowledgeBase',
            embeddings_model= bedrock.BedrockFoundationModel.TITAN_EMBED_TEXT_V1,
            instruction=  'Use this knowledge base to answer questions about books. ' +
    'It contains the full text of novels.'
        )

docBucket = s3.Bucket(self, 'DockBucket')

bedrock.S3DataSource(self, 'DataSource',
    bucket= docBucket,
    knowledge_base=kb,
    data_source_name='books',
    chunking_strategy= bedrock.ChunkingStrategy.FIXED_SIZE,
)

Example of Amazon RDS Aurora PostgreSQL:

TypeScript

import * as s3 from "aws-cdk-lib/aws-s3";
import { amazonaurora, bedrock } from "@cdklabs/generative-ai-cdk-constructs";

// Dimension of your vector embedding
embeddingsModelVectorDimension = 1024;
const auroraDb = new amazonaurora.AmazonAuroraVectorStore(stack, "AuroraDefaultVectorStore", {
  embeddingsModelVectorDimension: embeddingsModelVectorDimension,
});

const kb = new bedrock.KnowledgeBase(this, "KnowledgeBase", {
  vectorStore: auroraDb,
  embeddingsModelVectorDimension: embeddingsModelVectorDimension,
  instruction: "Use this knowledge base to answer questions about books. " + "It contains the full text of novels.",
});

const docBucket = new s3.Bucket(this, "DocBucket");

new bedrock.S3DataSource(this, "DataSource", {
  bucket: docBucket,
  knowledgeBase: kb,
  dataSourceName: "books",
  chunkingStrategy: bedrock.ChunkingStrategy.FIXED_SIZE,
});

Python

from aws_cdk import (
    aws_s3 as s3,
    aws_rds as rds,
    aws_ec2 as ec2,
    Stack,
    ArnFormat
)
from cdklabs.generative_ai_cdk_constructs import (
    bedrock,
    amazonaurora,
)

# Dimension of your vector embedding
embeddings_model_vector_dimension = 1024
aurora_db = amazonaurora.AmazonAuroraVectorStore(self, 'AuroraDefaultVectorStore',
  embeddings_model_vector_dimension=embeddings_model_vector_dimension
)

kb = bedrock.KnowledgeBase(self, 'KnowledgeBase',
            vector_store= aurora_db,
            embeddings_model_vector_dimension=embeddings_model_vector_dimension,
            instruction=  'Use this knowledge base to answer questions about books. ' +
    'It contains the full text of novels.'
        )

docBucket = s3.Bucket(self, 'DockBucket')

bedrock.S3DataSource(self, 'DataSource',
    bucket= docBucket,
    knowledge_base=kb,
    data_source_name='books',
    chunking_strategy= bedrock.ChunkingStrategy.FIXED_SIZE,
)

Example of importing existing Amazon RDS Aurora PostgreSQL using fromExistingAuroraVectorStore() method. Note - you need to provide clusterIdentifier, databaseName, vpc, secret and auroraSecurityGroupId used in deployment of your existing RDS Amazon Aurora DB, as well as embeddingsModel that you want to be used by a Knowledge Base for chunking:

TypeScript

import * as s3 from "aws-cdk-lib/aws-s3";
import { amazonaurora, bedrock } from '@cdklabs/generative-ai-cdk-constructs';

const auroraDb = aurora.AmazonAuroraVectorStore.fromExistingAuroraVectorStore(stack, 'ExistingAuroraVectorStore', {
  clusterIdentifier: 'aurora-serverless-vector-cluster',
  databaseName: 'bedrock_vector_db',
  schemaName: 'bedrock_integration',
  tableName: 'bedrock_kb',
  vectorField: 'embedding',
  textField: 'chunks',
  metadataField: 'metadata',
  primaryKeyField: 'id',
  embeddingsModel: bedrock.BedrockFoundationModel.COHERE_EMBED_ENGLISH_V3,
  vpc: cdk.aws_ec2.Vpc.fromLookup(stack, 'VPC', {
    vpcId: 'vpc-0c1a234567ee8bc90',
  }),
  auroraSecurityGroupId: 'sg-012ef345678c98a76',,
  secret: cdk.aws_rds.DatabaseSecret.fromSecretCompleteArn(
    stack,
    'Secret',
    cdk.Stack.of(stack).formatArn({
      service: 'secretsmanager',
      resource: 'secret',
      resourceName: 'rds-db-credentials/cluster-1234567890',
      region: cdk.Stack.of(stack).region,
      account: cdk.Stack.of(stack).account,
      arnFormat: cdk.ArnFormat.COLON_RESOURCE_NAME,
    }),
  ),
});

const kb = new bedrock.KnowledgeBase(this, "KnowledgeBase", {
  vectorStore: auroraDb,
  embeddingsModel: bedrock.BedrockFoundationModel.COHERE_EMBED_ENGLISH_V3,
  instruction:
    "Use this knowledge base to answer questions about books. " +
    "It contains the full text of novels.",
});

const docBucket = new s3.Bucket(this, "DocBucket");

new bedrock.S3DataSource(this, "DataSource", {
  bucket: docBucket,
  knowledgeBase: kb,
  dataSourceName: "books",
  chunkingStrategy: bedrock.ChunkingStrategy.FIXED_SIZE,
});

Python

from aws_cdk import (
    aws_s3 as s3,
    aws_rds as rds,
    aws_ec2 as ec2,
    Stack,
    ArnFormat
)
from cdklabs.generative_ai_cdk_constructs import (
    bedrock,
    amazonaurora,
)

aurora_db = amazonaurora.AmazonAuroraVectorStore.from_existing_aurora_vector_store(
    self, 'ExistingAuroraVectorStore',
    cluster_identifier='aurora-serverless-vector-cluster',
    database_name='bedrock_vector_db',
    schema_name='bedrock_integration',
    table_name='bedrock_kb',
    vector_field='embedding',
    text_field='chunks',
    metadata_field='metadata',
    primary_key_field='id',
    embeddings_model=bedrock.BedrockFoundationModel.COHERE_EMBED_ENGLISH_V3,
    vpc=ec2.Vpc.from_lookup(self, 'VPC', vpc_id='vpc-0c1a234567ee8bc90'),
    aurora_security_group_id='sg-012ef345678c98a76',,
    secret=rds.DatabaseSecret.from_secret_complete_arn(
        self,
        'Secret',
        Stack.of(self).format_arn(
            service= 'secretsmanager',
            resource= 'secret',
            resource_name= 'rds-db-credentials/cluster-1234567890',
            region= Stack.of(self).region,
            account= Stack.of(self).account,
            arn_format= ArnFormat.COLON_RESOURCE_NAME
        )
    )
)

kb = bedrock.KnowledgeBase(self, 'KnowledgeBase',
            vector_store= aurora_db,
            embeddings_model= bedrock.BedrockFoundationModel.COHERE_EMBED_ENGLISH_V3,
            instruction=  'Use this knowledge base to answer questions about books. ' +
    'It contains the full text of novels.'
        )

docBucket = s3.Bucket(self, 'DockBucket')

bedrock.S3DataSource(self, 'DataSource',
    bucket= docBucket,
    knowledge_base=kb,
    data_source_name='books',
    chunking_strategy= bedrock.ChunkingStrategy.FIXED_SIZE,
)

Example of Pinecone (manual, you must have Pinecone vector store created):

TypeScript

import * as s3 from "aws-cdk-lib/aws-s3";
import { pinecone, bedrock } from "@cdklabs/generative-ai-cdk-constructs";

const pineconeds = new pinecone.PineconeVectorStore({
  connectionString: "https://your-index-1234567.svc.gcp-starter.pinecone.io",
  credentialsSecretArn: "arn:aws:secretsmanager:your-region:123456789876:secret:your-key-name",
  textField: "question",
  metadataField: "metadata",
});

const kb = new bedrock.KnowledgeBase(this, "KnowledgeBase", {
  vectorStore: pineconeds,
  embeddingsModel: bedrock.BedrockFoundationModel.TITAN_EMBED_TEXT_V1,
  instruction: "Use this knowledge base to answer questions about books. " + "It contains the full text of novels.",
});

const docBucket = new s3.Bucket(this, "DocBucket");

new bedrock.S3DataSource(this, "DataSource", {
  bucket: docBucket,
  knowledgeBase: kb,
  dataSourceName: "books",
  chunkingStrategy: bedrock.ChunkingStrategy.FIXED_SIZE,
});

Python

from aws_cdk import (
    aws_s3 as s3,
)
from cdklabs.generative_ai_cdk_constructs import (
    bedrock,
    pinecone,
)

pineconevs = pinecone.PineconeVectorStore(
            connection_string='https://your-index-1234567.svc.gcp-starter.pinecone.io',
            credentials_secret_arn='arn:aws:secretsmanager:your-region:123456789876:secret:your-key-name',
            text_field='question',
            metadata_field='metadata'
        )

kb = bedrock.KnowledgeBase(self, 'KnowledgeBase',
            vector_store= pineconevs,
            embeddings_model= bedrock.BedrockFoundationModel.COHERE_EMBED_ENGLISH_V3,
            instruction=  'Use this knowledge base to answer questions about books. ' +
    'It contains the full text of novels.'
        )

docBucket = s3.Bucket(self, 'DockBucket')

bedrock.S3DataSource(self, 'DataSource',
    bucket= docBucket,
    knowledge_base=kb,
    data_source_name='books',
    chunking_strategy= bedrock.ChunkingStrategy.FIXED_SIZE,
)

Knowledge Base - Data Sources

Data sources are the various repositories or systems from which information is extracted and ingested into the knowledge base. These sources provide the raw content that will be processed, indexed, and made available for querying within the knowledge base system. Data sources can include various types of systems such as document management systems, databases, file storage systems, and content management platforms. Suuported Data Sources include Amazon S3 buckets, Web Crawlers, SharePoint sites, Salesforce instances, and Confluence spaces.

  • Amazon S3. You can either create a new data source using the bedrock.S3DataSource(..) class, or using the kb.addS3DataSource(..).
  • Web Crawler. You can either create a new data source using the bedrock.WebCrawlerDataSource(..) class, or using the kb.addWebCrawlerDataSource(..).
  • Confluence. You can either create a new data source using the bedrock.ConfluenceDataSource(..) class, or using the kb.addConfluenceDataSource(..).
  • SharePoint. You can either create a new data source using the bedrock.SharePointDataSource(..) class, or using the kb.addSharePointDataSource(..).
  • Salesforce. You can either create a new data source using the bedrock.SalesforceDataSource(..) class, or using the kb.addSalesforceDataSource(..).

Typescript

const app = new cdk.App();
const stack = new cdk.Stack(app, "aws-cdk-bedrock-data-sources-integ-test");

const kb = new KnowledgeBase(stack, "MyKnowledgeBase", {
  name: "MyKnowledgeBase",
  embeddingsModel: BedrockFoundationModel.COHERE_EMBED_MULTILINGUAL_V3,
});

const bucket = new Bucket(stack, "Bucket", {});
const lambdaFunction = new Function(stack, "MyFunction", {
  runtime: cdk.aws_lambda.Runtime.PYTHON_3_9,
  handler: "index.handler",
  code: cdk.aws_lambda.Code.fromInline('print("Hello, World!")'),
});

const secret = new Secret(stack, "Secret");
const key = new Key(stack, "Key");

kb.addWebCrawlerDataSource({
  sourceUrls: ["https://docs.aws.amazon.com/"],
  chunkingStrategy: ChunkingStrategy.HIERARCHICAL_COHERE,
  customTransformation: CustomTransformation.lambda({
    lambdaFunction: lambdaFunction,
    s3BucketUri: `s3://${bucket.bucketName}/chunk-processor/`,
  }),
});

kb.addS3DataSource({
  bucket,
  chunkingStrategy: ChunkingStrategy.SEMANTIC,
  parsingStrategy: ParsingStategy.foundationModel({
    model: BedrockFoundationModel.ANTHROPIC_CLAUDE_SONNET_V1_0,
  }),
});

kb.addConfluenceDataSource({
  dataSourceName: "TestDataSource",
  authSecret: secret,
  kmsKey: key,
  confluenceUrl: "https://example.atlassian.net",
  filters: [
    {
      objectType: ConfluenceObjectType.ATTACHMENT,
      includePatterns: [".*\\.pdf"],
      excludePatterns: [".*private.*\\.pdf"],
    },
    {
      objectType: ConfluenceObjectType.PAGE,
      includePatterns: [".*public.*\\.pdf"],
      excludePatterns: [".*confidential.*\\.pdf"],
    },
  ],
});

kb.addSalesforceDataSource({
  authSecret: secret,
  endpoint: "https://your-instance.my.salesforce.com",
  kmsKey: key,
  filters: [
    {
      objectType: SalesforceObjectType.ATTACHMENT,
      includePatterns: [".*\\.pdf"],
      excludePatterns: [".*private.*\\.pdf"],
    },
    {
      objectType: SalesforceObjectType.CONTRACT,
      includePatterns: [".*public.*\\.pdf"],
      excludePatterns: [".*confidential.*\\.pdf"],
    },
  ],
});

kb.addSharePointDataSource({
  dataSourceName: "SharepointDataSource",
  authSecret: secret,
  kmsKey: key,
  domain: "yourdomain",
  siteUrls: ["https://yourdomain.sharepoint.com/sites/mysite"],
  tenantId: "888d0b57-69f1-4fb8-957f-e1f0bedf64de",
  filters: [
    {
      objectType: SharePointObjectType.PAGE,
      includePatterns: [".*\\.pdf"],
      excludePatterns: [".*private.*\\.pdf"],
    },
    {
      objectType: SharePointObjectType.FILE,
      includePatterns: [".*public.*\\.pdf"],
      excludePatterns: [".*confidential.*\\.pdf"],
    },
  ],
});

Python

from aws_cdk import (
    Stack,
    aws_s3 as s3,
    aws_lambda as _lambda,
    aws_secretsmanager as secretsmanager,
    aws_kms as kms
)
from constructs import Construct
from cdklabs.generative_ai_cdk_constructs import (
    bedrock
)

class PythonTestStack(Stack):

    def __init__(self, scope: Construct, construct_id: str, **kwargs) -> None:
        super().__init__(scope, construct_id, **kwargs)

        kb = bedrock.KnowledgeBase(self, 'MyKnowledgeBase',
                    embeddings_model= bedrock.BedrockFoundationModel.COHERE_EMBED_MULTILINGUAL_V3,
                )

        docBucket = s3.Bucket(self, 'Bucket')

        function = _lambda.Function(self, 'MyFunction',
            runtime=_lambda.Runtime.PYTHON_3_12,
            handler='index.handler',
            code=_lambda.Code.from_inline('print("Hello, World!")'),
        )

        kb.add_web_crawler_data_source(
            source_urls= ['https://docs.aws.amazon.com/'],
            chunking_strategy= bedrock.ChunkingStrategy.HIERARCHICAL_COHERE,
            custom_transformation= bedrock.CustomTransformation.lambda_(
                lambda_function= function,
                s3_bucket_uri= f's3://{docBucket.bucket_name}/chunk-processor/'
            )
        )

        kb.add_s3_data_source(
            bucket= docBucket,
            chunking_strategy= bedrock.ChunkingStrategy.SEMANTIC,
            parsing_strategy= bedrock.ParsingStategy.foundation_model(
                parsing_model= bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_3_5_SONNET_V1_0.as_i_model(self)
            )
        )

        secret = secretsmanager.Secret(self, 'Secret')
        key = kms.Key(self, 'Key')

        kb.add_confluence_data_source(
            data_source_name='TestDataSource',
            auth_secret=secret,
            kms_key=key,
            confluence_url='https://example.atlassian.net',
            filters=[
                bedrock.ConfluenceCrawlingFilters(
                    object_type=bedrock.ConfluenceObjectType.ATTACHMENT,
                    include_patterns= [".*\\.pdf"],
                    exclude_patterns= [".*private.*\\.pdf"],
                ),
                bedrock.ConfluenceCrawlingFilters(
                    object_type=bedrock.ConfluenceObjectType.PAGE,
                    include_patterns= [".*public.*\\.pdf"],
                    exclude_patterns= [".*confidential.*\\.pdf"],
                ),
            ]
        )

        kb.add_salesforce_data_source(
            auth_secret=secret,
            endpoint='https://your-instance.my.salesforce.com',
            kms_key=key,
            filters=[
                bedrock.SalesforceCrawlingFilters(
                    object_type=bedrock.SalesforceObjectType.ATTACHMENT,
                    include_patterns= [".*\\.pdf"],
                    exclude_patterns= [".*private.*\\.pdf"],
                ),
                bedrock.SalesforceCrawlingFilters(
                    object_type=bedrock.SalesforceObjectType.CONTRACT,
                    include_patterns= [".*public.*\\.pdf"],
                    exclude_patterns= [".*confidential.*\\.pdf"],
                ),
            ]
        )

        kb.add_share_point_data_source(
            data_source_name='SharepointDataSource',
            auth_secret=secret,
            kms_key=key,
            domain='yourDomain',
            site_urls= ['https://yourdomain.sharepoint.com/sites/mysite'],
            tenant_id='888d0b57-69f1-4fb8-957f-e1f0bedf64de',
            filters=[
                bedrock.SharePointCrawlingFilters(
                    object_type=bedrock.SharePointObjectType.PAGE,
                    include_patterns= [".*\\.pdf"],
                    exclude_patterns= [".*private.*\\.pdf"],
                ),
                bedrock.SharePointCrawlingFilters(
                    object_type=bedrock.SharePointObjectType.FILE,
                    include_patterns= [".*public.*\\.pdf"],
                    exclude_patterns= [".*confidential.*\\.pdf"],
                ),
            ]
        )

Knowledge Base - Chunking Strategies

  • Default Chunking: Applies Fixed Chunking with the default chunk size of 300 tokens and 20% overlap.

    TypeScript

    ChunkingStrategy.DEFAULT;

    Python

    ChunkingStrategy.DEFAULT;

  • Fixed Size Chunking: This method divides the data into fixed-size chunks, with each chunk containing a predetermined number of tokens. This strategy is useful when the data is uniform in size and structure. Typescript

    TypeScript

    // Fixed Size Chunking with sane defaults.
ChunkingStrategy.FIXED_SIZE;

// Fixed Size Chunking with custom values.
ChunkingStrategy.fixedSize({ maxTokens: 200, overlapPercentage: 25 });

    Python

    # Fixed Size Chunking with sane defaults.
ChunkingStrategy.FIXED_SIZE;

# Fixed Size Chunking with custom values.
ChunkingStrategy.fixed_size(
  max_tokens= 200,
  overlap_percentage= 25
)

  • Hierarchical Chunking: This strategy organizes data into layers of chunks, with the first layer containing large chunks and the second layer containing smaller chunks derived from the first. It is ideal for data with inherent hierarchies or nested structures.

    TypeScript

    // Hierarchical Chunking with the default for Cohere Models.
ChunkingStrategy.HIERARCHICAL_COHERE;

// Hierarchical Chunking with the default for Titan Models.
ChunkingStrategy.HIERARCHICAL_TITAN;

// Hierarchical Chunking with custom values. Tthe maximum chunk size depends on the model.
// Amazon Titan Text Embeddings: 8192. Cohere Embed models: 512
ChunkingStrategy.hierarchical({
  overlapTokens: 60,
  maxParentTokenSize: 1500,
  maxChildTokenSize: 300,
});

    Python

    # Hierarchical Chunking with the default for Cohere Models.
ChunkingStrategy.HIERARCHICAL_COHERE

# Hierarchical Chunking with the default for Titan Models.
ChunkingStrategy.HIERARCHICAL_TITAN

# Hierarchical Chunking with custom values. Tthe maximum chunk size depends on the model.
# Amazon Titan Text Embeddings: 8192. Cohere Embed models: 512
chunking_strategy= ChunkingStrategy.hierarchical(
    overlap_tokens=60,
    max_parent_token_size=1500,
    max_child_token_size=300
)

  • Semantic Chunking: This method splits data into smaller documents based on groups of similar content derived from the text using natural language processing. It helps preserve contextual relationships and ensures accurate and contextually appropriate results.

    TypeScript

    // Semantic Chunking with sane defaults.
ChunkingStrategy.SEMANTIC;

// Semantic Chunking with custom values.
ChunkingStrategy.semantic({ bufferSize: 0, breakpointPercentileThreshold: 95, maxTokens: 300 });

    Python

    # Semantic Chunking with sane defaults.
ChunkingStrategy.SEMANTIC

# Semantic Chunking with custom values.
ChunkingStrategy.semantic(
  buffer_size=0,
  breakpoint_percentile_threshold=95,
  max_tokens=300
)

  • No Chunking: This strategy treats each file as one chunk. If you choose this option, you may want to pre-process your documents by splitting them into separate files.

    TypeScript

    ChunkingStrategy.NONE;

    Python

    ChunkingStrategy.NONE;

Knowledge Base - Parsing Strategy

A parsing strategy in Amazon Bedrock is a configuration that determines how the service processes and interprets the contents of a document. It involves converting the document's contents into text and splitting it into smaller chunks for analysis. Amazon Bedrock offers two parsing strategies:

  • Default Parsing Strategy: This strategy converts the document's contents into text and splits it into chunks using a predefined approach. It is suitable for most use cases but may not be optimal for specific document types or requirements.

  • Foundation Model Parsing Strategy: This strategy uses a foundation model to describe the contents of the document. It is particularly useful for improved processing of PDF files with tables and images. To use this strategy, set the parsingStrategy in a data source as below.

    TypeScript

    bedrock.ParsingStategy.foundationModel({
  model: BedrockFoundationModel.ANTHROPIC_CLAUDE_SONNET_V1_0,
});

    Python

    bedrock.ParsingStategy.foundation_model(
    parsing_model=BedrockFoundationModel.ANTHROPIC_CLAUDE_SONNET_V1_0
)

Knowledge Base - Custom Transformation

Custom Transformation in Amazon Bedrock is a feature that allows you to create and apply custom processing steps to documents moving through a data source ingestion pipeline.

Custom Transformation uses AWS Lambda functions to process documents, enabling you to perform custom operations such as data extraction, normalization, or enrichment. To create a custom transformation, set the customTransformation in a data source as below.

TypeScript

CustomTransformation.lambda({
lambdaFunction: lambdaFunction,
s3BucketUri: `s3://${bucket.bucketName}/chunk-processor/`,
}),

Python

CustomTransformation.lambda_(
  lambda_function= function,
  s3_bucket_uri= f's3://{docBucket.bucket_name}/chunk-processor/'
)

Agents

Enable generative AI applications to execute multistep tasks across company systems and data sources.

Create an Agent

The following example creates an Agent with a simple instruction and default prompts that consults a Knowledge Base.

TypeScript

const agent = new bedrock.Agent(this, "Agent", {
  foundationModel: bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_V2_1,
  instruction: "You are a helpful and friendly agent that answers questions about literature.",
});

agent.addKnowledgeBase(kb);

Python

agent = bedrock.Agent(
    self,
    "Agent",
    foundation_model=bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_V2_1,
    instruction="You are a helpful and friendly agent that answers questions about insurance claims.",
)
  agent.add_knowledge_base(kb);

You can also use system defined inference profiles to enable cross region inference requests for supported models. For instance:

TypeScript

const cris = bedrock.CrossRegionInferenceProfile.fromConfig({
  geoRegion: bedrock.CrossRegionInferenceProfileRegion.US,
  model: bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_3_5_SONNET_V1_0,
});

const agent = new bedrock.Agent(this, 'Agent', {
  foundationModel: cris,
  instruction: 'You are a helpful and friendly agent that answers questions about agriculture.'
});

Python

cris = bedrock.CrossRegionInferenceProfile.from_config(
  geo_region= bedrock.CrossRegionInferenceProfileRegion.US,
  model= bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_3_5_SONNET_V1_0
)

agent = bedrock.Agent(
    self,
    "Agent",
    foundation_model=cris,
    instruction="You are a helpful and friendly agent that answers questions about agriculture.",
)

For more information on cross region inference, please refer to System defined inference profiles

Action Groups

An action group defines functions your agent can call. The functions are Lambda functions. The action group uses an OpenAPI schema to tell the agent what your functions do and how to call them.

const actionGroupFunction = new lambda_python.PythonFunction(this, "ActionGroupFunction", {
  runtime: lambda.Runtime.PYTHON_3_12,
  entry: path.join(__dirname, "../lambda/action-group"),
});

const actionGroup = new bedrock.AgentActionGroup(this, "MyActionGroup", {
  actionGroupName: "query-library",
  description: "Use these functions to get information about the books in the library.",
  actionGroupExecutor: {
    lambda: actionGroupFunction,
  },
  actionGroupState: "ENABLED",
  apiSchema: bedrock.ApiSchema.fromAsset(path.join(__dirname, "action-group.yaml")),
});

agent.addActionGroup(actionGroup);

Python

action_group_function = PythonFunction(
            self,
            "LambdaFunction",
            runtime=Runtime.PYTHON_3_12,
            entry="./lambda",
            index="app.py",
            handler="lambda_handler",
)

actionGroup = bedrock.AgentActionGroup(self,
    "MyActionGroup",
    action_group_name="query-library",
    description="Use these functions to get information about the books in the library.",
    action_group_executor= bedrock.ActionGroupExecutor(
      lambda_=action_group_function
    ),
    action_group_state="ENABLED",
    api_schema=bedrock.ApiSchema.from_asset("action-group.yaml"))

agent.add_action_group(actionGroup)

Prepare the Agent

The Agent constructs take an optional parameter shouldPrepareAgent to indicate that the Agent should be prepared after any updates to an agent, Knowledge Base association, or action group. This may increase the time to create and update those resources. By default, this value is false .

Creating an agent alias will not prepare the agent, so if you create an alias with addAlias or by providing an aliasName when creating the agent then you should set shouldPrepareAgent to true.

Prompt Overrides

Bedrock Agents allows you to customize the prompts and LLM configuration for its different steps. You can disable steps or create a new prompt template. Prompt templates can be inserted from plain text files.

TypeScript

import { readFileSync } from "fs";

const orchestration = readFileSync("prompts/orchestration.txt", "utf-8");
const agent = new bedrock.Agent(this, "Agent", {
  foundationModel: bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_V2_1,
  instruction: "You are a helpful and friendly agent that answers questions about literature.",
  promptOverrideConfiguration: {
    promptConfigurations: [
      {
        promptType: bedrock.PromptType.PRE_PROCESSING,
        promptState: bedrock.PromptState.DISABLED,
        promptCreationMode: bedrock.PromptCreationMode.OVERRIDDEN,
        basePromptTemplate: "disabled",
        inferenceConfiguration: {
          temperature: 0.0,
          topP: 1,
          topK: 250,
          maximumLength: 1,
          stopSequences: ["\n\nHuman:"],
        },
      },
      {
        promptType: bedrock.PromptType.ORCHESTRATION,
        basePromptTemplate: orchestration,
        promptState: bedrock.PromptState.ENABLED,
        promptCreationMode: bedrock.PromptCreationMode.OVERRIDDEN,
        inferenceConfiguration: {
          temperature: 0.0,
          topP: 1,
          topK: 250,
          maximumLength: 2048,
          stopSequences: ["</invoke>", "</answer>", "</error>"],
        },
      },
    ],
  },
});

Python

orchestration = open('prompts/orchestration.txt', encoding="utf-8").read()
agent = bedrock.Agent(self, "Agent",
            foundation_model=bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_V2_1,
            instruction="You are a helpful and friendly agent that answers questions about insurance claims.",
            prompt_override_configuration= bedrock.PromptOverrideConfiguration(
                prompt_configurations=[
                    bedrock.PromptConfiguration(
                        prompt_type=bedrock.PromptType.PRE_PROCESSING,
                        prompt_state=bedrock.PromptState.DISABLED,
                        prompt_creation_mode=bedrock.PromptCreationMode.OVERRIDDEN,
                        base_prompt_template="disabled",
                        inference_configuration=bedrock.InferenceConfiguration(
                            temperature=0.0,
                            top_k=250,
                            top_p=1,
                            maximum_length=1,
                            stop_sequences=['\n\nHuman:'],
                        )
                    ),
                    bedrock.PromptConfiguration(
                        prompt_type=bedrock.PromptType.ORCHESTRATION,
                        prompt_state=bedrock.PromptState.ENABLED,
                        prompt_creation_mode=bedrock.PromptCreationMode.OVERRIDDEN,
                        base_prompt_template=orchestration,
                        inference_configuration=bedrock.InferenceConfiguration(
                            temperature=0.0,
                            top_k=250,
                            top_p=1,
                            maximum_length=2048,
                            stop_sequences=['</invoke>', '</answer>', '</error>'],
                        )
                    )
                ]
            ),
        )

Agent Alias

After you have sufficiently iterated on your working draft and are satisfied with the behavior of your agent, you can set it up for deployment and integration into your application by creating aliases of your agent.

To deploy your agent, you need to create an alias. During alias creation, Amazon Bedrock automatically creates a version of your agent. The alias points to this newly created version. You can point the alias to a previously created version if necessary. You then configure your application to make API calls to that alias.

By default, the Agent resource does not create any aliases, and you can use the 'DRAFT' version.

Tracking the latest version

The Agent resource optionally takes an aliasName property that, if defined, will create an Alias that creates a new version on every change.

TypeScript

const agent = new bedrock.Agent(this, "Agent", {
  foundationModel: bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_V2_1,
  instruction: "You are a helpful and friendly agent that answers questions about literature.",
  knowledgeBases: [kb],
  aliasName: "latest",
});

Python

agent = bedrock.Agent(
    self,
    "Agent",
    foundation_model=bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_V2_1,
    instruction="You are a helpful and friendly agent that answers questions about insurance claims.",
    knowledge_bases= [kb],
    alias_name='latest'
)

Specific version

Using the addAlias method you can create aliases with a specific agent version.

TypeScript

agent.addAlias({
  aliasName: "prod",
  agentVersion: "12",
});

Python

agent.add_alias(
    alias_name='prod',
    agent_version='12'
)

Alternatively, you can use the AgentAlias resource if you want to create an Alias for an existing Agent.

TypeScript

const alias = new bedrock.AgentAlias(this, "ProdAlias", {
  agentId: "ABCDE12345",
  aliasName: "prod",
  agentVersion: "12",
});

Python

alias = bedrock.AgentAlias(self, 'ProdAlias',
    agent_id='ABCDE12345',
    alias_name='prod',
    agent_version='12'
)

Bedrock Guardrails

Amazon Bedrock's Guardrails feature enables you to implement robust governance and control mechanisms for your generative AI applications, ensuring alignment with your specific use cases and responsible AI policies. Guardrails empowers you to create multiple tailored policy configurations, each designed to address the unique requirements and constraints of different use cases. These policy configurations can then be seamlessly applied across multiple foundation models (FMs) and Agents, ensuring a consistent user experience and standardizing safety, security, and privacy controls throughout your generative AI ecosystem.

With Guardrails, you can define and enforce granular, customizable policies to precisely govern the behavior of your generative AI applications. You can configure the following policies in a guardrail to avoid undesirable and harmful content and remove sensitive information for privacy protection.

Content filters – Adjust filter strengths to block input prompts or model responses containing harmful content.

Denied topics – Define a set of topics that are undesirable in the context of your application. These topics will be blocked if detected in user queries or model responses.

Word filters – Configure filters to block undesirable words, phrases, and profanity. Such words can include offensive terms, competitor names etc.

Sensitive information filters – Block or mask sensitive information such as personally identifiable information (PII) or custom regex in user inputs and model responses.

You can create a Guardrail with a minimum blockedInputMessaging ,blockedOutputsMessaging and default content filter policy.

TypeScript

const guardrails = new bedrock.Guardrail(this, "bedrockGuardrails", {
  name: "my-BedrockGuardrails",
  description: "Legal ethical guardrails.",
});

// Optional - Add Sensitive information filters

guardrail.addPIIFilter({
  type: PIIType.General.ADDRESS,
  action: GuardrailAction.ANONYMIZE,
});

guardrail.addRegexFilter({
  name: "TestRegexFilter",
  description: "This is a test regex filter",
  pattern: "/^[A-Z]{2}d{6}$/",
  action: bedrock.GuardrailAction.ANONYMIZE,
});

// Optional - Add contextual grounding

guardrail.addContextualGroundingFilter({
  type: ContextualGroundingFilterType.GROUNDING,
  threshold: 0.95,
});

guardrail.addContextualGroundingFilter({
  type: ContextualGroundingFilterType.RELEVANCE,
  threshold: 0.95,
});

// Optional - Add Denied topics . You can use a Topic or create your custom Topic

guardrail.addDeniedTopicFilter(Topic.FINANCIAL_ADVICE);
guardrail.addDeniedTopicFilter(
  Topic.custom({
    name: "Legal_Advice",
    definition:
      "Offering guidance or suggestions on legal matters, legal actions, interpretation of laws, or legal rights and responsibilities.",
    examples: [
      "Can I sue someone for this?",
      "What are my legal rights in this situation?",
      "Is this action against the law?",
      "What should I do to file a legal complaint?",
      "Can you explain this law to me?",
    ],
  })
);

// Optional - Add Word filters. You can upload words from a file with addWordFilterFromFile function.
guardrail.addWordFilter("drugs");
guardrail.addManagedWordListFilter(ManagedWordFilterType.PROFANITY);
guardrails.addWordFilterFromFile("./scripts/wordsPolicy.csv");

// versioning - if you change any guardrail configuration, a new version will be created
guardrails.createVersion("testversion");

// Importing existing guardrail
const importedGuardrail = bedrock.Guardrail.fromGuardrailAttributes(stack, "TestGuardrail", {
  guardrailArn: "arn:aws:bedrock:us-east-1:123456789012:guardrail/oygh3o8g7rtl",
  guardrailVersion: "1", //optional
  kmsKey: kmsKey, //optional
});

// Importing Guardrails created through the L1 CDK CfnGuardrail construct
const cfnGuardrail = new CfnGuardrail(this, "MyCfnGuardrail", {
  blockedInputMessaging: "blockedInputMessaging",
  blockedOutputsMessaging: "blockedOutputsMessaging",
  name: "namemycfnguardrails",
  wordPolicyConfig: {
    wordsConfig: [
      {
        text: "drugs",
      },
    ],
  },
});

const importedGuardrail = bedrock.Guardrail.fromCfnGuardrail(cfnGuardrail);

Python

    guardrail = bedrock.Guardrail(self, 'myGuardrails',
        name='my-BedrockGuardrails',
        description= "Legal ethical guardrails.")

    # Optional - Add Sensitive information filters

    guardrail.add_pii_filter(
        type= bedrock.pii_type.General.ADDRESS,
        action= bedrock.GuardrailAction.ANONYMIZE,
    )

    guardrail.add_regex_filter(
        name= "TestRegexFilter",
        description= "This is a test regex filter",
        pattern= "/^[A-Z]{2}d{6}$/",
        action= bedrock.GuardrailAction.ANONYMIZE,
    )

    # Optional - Add contextual grounding

    guardrail.add_contextual_grounding_filter(
        type= bedrock.ContextualGroundingFilterType.GROUNDING,
        threshold= 0.95,
    )

    # Optional - Add Denied topics . You can use default Topic or create your custom Topic with createTopic function. The default Topics can also be overwritten.

    guardrail.add_contextual_grounding_filter(
        type= bedrock.ContextualGroundingFilterType.RELEVANCE,
        threshold= 0.95,
    )

    guardrail.add_denied_topic_filter(bedrock.Topic.FINANCIAL_ADVICE)

    guardrail.add_denied_topic_filter(
      bedrock.Topic.custom(
        name= "Legal_Advice",
        definition=
            "Offering guidance or suggestions on legal matters, legal actions, interpretation of laws, or legal rights and responsibilities.",
        examples= [
            "Can I sue someone for this?",
            "What are my legal rights in this situation?",
            "Is this action against the law?",
            "What should I do to file a legal complaint?",
            "Can you explain this law to me?",
        ]
      )
    )

    # Optional - Add Word filters. You can upload words from a file with addWordFilterFromFile function.
    guardrail.add_word_filter("drugs")
    guardrail.add_managed_word_list_filter(bedrock.ManagedWordFilterType.PROFANITY)
    guardrail.add_word_filter_from_file("./scripts/wordsPolicy.csv")

    # versioning - if you change any guardrail configuration, a new version will be created
    guardrail.create_version("testversion")

    # Importing existing guardrail
    imported_guardrail = bedrock.Guardrail.from_guardrail_attributes(self, "TestGuardrail",
      guardrail_arn="arn:aws:bedrock:us-east-1:123456789012:guardrail/oygh3o8g7rtl",
      guardrail_version="1",
      kms_key=kms_key
    )

    # Importing Guardrails created through the L1 CDK CfnGuardrail construct
    cfn_guardrail = cfnbedrock.CfnGuardrail(self, "MyCfnGuardrail",
        blocked_input_messaging="blockedInputMessaging",
        blocked_outputs_messaging="blockedOutputsMessaging",
        name="name",

        # the properties below are optional
        word_policy_config=cfnbedrock.CfnGuardrail.WordPolicyConfigProperty(
            words_config=[cfnbedrock.CfnGuardrail.WordConfigProperty(
                text="drugs"
            )]
        )
    )

    imported_guardrail = bedrock.Guardrail.from_cfn_guardrail(cfn_guardrail)

Prompt management

Amazon Bedrock provides the ability to create and save prompts using Prompt management so that you can save time by applying the same prompt to different workflows. You can include variables in the prompt so that you can adjust the prompt for different use case.

The Prompt resource allows you to create a new prompt. Example of Prompt:

TypeScript

const cmk = new kms.Key(this, "cmk", {});
const claudeModel = BedrockFoundationModel.ANTHROPIC_CLAUDE_SONNET_V1_0;

const variant1 = PromptVariant.text({
  variantName: "variant1",
  model: claudeModel,
  promptVariables: ["topic"],
  promptText: "This is my first text prompt. Please summarize our conversation on: {{topic}}.",
  inferenceConfiguration: {
    temperature: 1.0,
    topP: 0.999,
    maxTokens: 2000,
  },
});

const prompt1 = new Prompt(this, "prompt1", {
  promptName: "prompt1",
  description: "my first prompt",
  defaultVariant: variant1,
  variants: [variant1],
  encryptionKey: cmk,
});

Prompt Variants

Prompt variants in the context of Amazon Bedrock refer to alternative configurations of a prompt, including its message or the model and inference configurations used. Prompt variants allow you to create different versions of a prompt, test them, and save the variant that works best for your use case. You can add prompt variants to a prompt by creating a PromptVariant object and specify the variants on prompt creation, or by using the .addVariant(..) method on a Prompt object.

Example of PromptVariant:

TypeScript

...

const variant2 = PromptVariant.text({
  variantName: "variant2",
  model: claudeModel,
  promptVariables: [ "topic" ],
  promptText: "This is my second text prompt. Please summarize our conversation on: {{topic}}.",
  inferenceConfiguration: {
    temperature: 0.5,
    topP: 0.999,
    maxTokens: 2000,
  },
});

prompt1.addVariant(variant2);

Prompt Version

A prompt version is a snapshot of a prompt at a specific point in time that you create when you are satisfied with a set of configurations. Versions allow you to deploy your prompt and easily switch between different configurations for your prompt and update your application with the most appropriate version for your use-case.

You can create a Prompt version by using the PromptVersion class or by using the .createVersion(..) on a Prompt object. It is recommended to use the .createVersion(..) method. It uses a hash based mechanism to update the version whenever a certain configuration property changes.

TypeScript

new PromptVersion(prompt1, "my first version");

or alternatively:

prompt1.createVersion("my first version");

System defined inference profiles

You can build a CrossRegionInferenceProfile using a system defined inference profile. The inference profile will route requests to the Regions defined in the cross region (system-defined) inference profile that you choose. You can find the system defined inference profiles by navigating to your console (Amazon Bedrock -> Cross-region inference) or programmatically, for instance using boto3.

Before using creating a CrossRegionInferenceProfile, ensure that you have access to the models and regions defined in the inference profiles. For instance, if you see the system defined inference profile "us.anthropic.claude-3-5-sonnet-20241022-v2:0" defined in your region, the table mentions that inference requests will be routed to US East (Virginia) us-east-1, US East (Ohio) us-east-2 and US West (Oregon) us-west-2. Thus, you need to have model access enabled in those regions for the model anthropic.claude-3-5-sonnet-20241022-v2:0. You can then create the CrossRegionInferenceProfile as follows:

TypeScript

const cris = bedrock.CrossRegionInferenceProfile.fromConfig({
  geoRegion: bedrock.CrossRegionInferenceProfileRegion.US,
  model: bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_3_5_SONNET_V2_0
});

Python

cris = bedrock.CrossRegionInferenceProfile.from_config(
  geo_region= bedrock.CrossRegionInferenceProfileRegion.US,
  model= bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_3_5_SONNET_V2_0
)

Application inference profile

You can create an application inference profile with one or more Regions to track usage and costs when invoking a model.

To create an application inference profile for one Region, specify a foundation model. Usage and costs for requests made to that Region with that model will be tracked.

To create an application inference profile for multiple Regions, specify a cross region (system-defined) inference profile. The inference profile will route requests to the Regions defined in the cross region (system-defined) inference profile that you choose. Usage and costs for requests made to the Regions in the inference profile will be tracked. You can find the system defined inference profiles by navigating to your console (Amazon Bedrock -> Cross-region inference) or programmatically, for instance using boto3:

bedrock = session.client("bedrock", region_name="us-east-1")
bedrock.list_inference_profiles(typeEquals='SYSTEM_DEFINED')

Before using application inference profiles, ensure that:

  • You have appropriate IAM permissions
  • You have access to the models and regions defined in the inference profiles
  • Ensure proper configuration of the required API permissions for inference profile-related actions

Specifically the role you are assuming needs to have permissions for following actions in the IAM policy

"Action": [
      "bedrock:GetInferenceProfile",
      "bedrock:ListInferenceProfiles",
      "bedrock:DeleteInferenceProfile"
      "bedrock:TagResource",
      "bedrock:UntagResource",
      "bedrock:ListTagsForResource"
  ]

You can restrict to specific resources by applying "Resources" tag in the IAM policy.

"Resource": ["arn:aws:bedrock:*:*:application-inference-profile/*"]

TypeScript

// Create an application inference profile for one Region
// You can use the 'bedrock.BedrockFoundationModel' or pass the arn as a string
const appInfProfile1 = new ApplicationInferenceProfile(this, 'myapplicationprofile', {
  inferenceProfileName: 'claude 3 sonnet v1',
  modelSource: bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_SONNET_V1_0,
  tags: [{key: 'test', value: 'test'}]
});

// To create an application inference profile across regions, specify the cross region inference profile
const cris = bedrock.CrossRegionInferenceProfile.fromConfig({
  geoRegion: bedrock.CrossRegionInferenceProfileRegion.US,
  model: bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_3_5_SONNET_V2_0
});

const appInfProfile2 = new ApplicationInferenceProfile(this, 'myapplicationprofile2', {
  inferenceProfileName: 'claude 3 sonnet v1',
  modelSource: cris
});

// Import a Cfn L1 construct created application inference profile
const cfnapp = new CfnApplicationInferenceProfile(this, 'mytestaip3', {
  inferenceProfileName: 'mytest',
  modelSource: {
    copyFrom: 'arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-3-sonnet-20240229-v1:0',
  },
});

const appInfProfile3 = bedrock.ApplicationInferenceProfile.fromCfnApplicationInferenceProfile(cfnapp);

// Import an inference profile through attributes
const appInfProfile4 = bedrock.ApplicationInferenceProfile.fromApplicationInferenceProfileAttributes(this, 'TestAIP', {
  inferenceProfileArn: 'arn:aws:bedrock:us-east-1:XXXXX:application-inference-profile/ID',
  inferenceProfileIdentifier: 'arn:aws:bedrock:us-east-1:XXXXXXX:application-inference-profile/ID',
});

Python

# Create an application inference profile for one Region
# You can use the 'bedrock.BedrockFoundationModel' or pass the arn as a string
appInfProfile1 = bedrock.ApplicationInferenceProfile(self, 'myapplicationprofile',
  inference_profile_name='claude 3 sonnet v1',
  model_source=bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_SONNET_V1_0,
  tags=[CfnTag(
    key="key",
    value="value"
  )] 
)

# To create an application inference profile across regions, specify the cross region inference profile
cris = bedrock.CrossRegionInferenceProfile.from_config(
  geo_region= bedrock.CrossRegionInferenceProfileRegion.US,
  model= bedrock.BedrockFoundationModel.ANTHROPIC_CLAUDE_3_5_SONNET_V2_0
)

appInfProfile2 = bedrock.ApplicationInferenceProfile(self, 'myapplicationprofile2',
  inference_profile_name='claude 35 sonnet v2',
  model_source=cris
)

# Import an inference profile through attributes
appInfProfile3 = bedrock.ApplicationInferenceProfile.from_application_inference_profile_attributes(self, 'TestAIP',
  inference_profile_arn='arn:aws:bedrock:us-east-1:XXXXX:application-inference-profile/ID',
  inference_profile_identifier='arn:aws:bedrock:us-east-1:XXXXXXX:application-inference-profile/ID',
)

# Import a Cfn L1 construct created application inference profile
cfnaip = CfnApplicationInferenceProfile(this, 'mytestaip4',
  inference_profile_name='mytest',
  model_source= CfnApplicationInferenceProfile.InferenceProfileModelSourceProperty(
    copy_from='arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-3-sonnet-20240229-v1:0'
  ),
)

appInfProfile4 = bedrock.ApplicationInferenceProfile.from_cfn_application_inference_profile(cfnaip);