AWS Bedrock Multi-Agent Blueprint with CDK & Streamlit

The AWS Bedrock Multi-Agent Blueprint is a project that allows you to create and deploy multi-agent applications on AWS Bedrock. It is built with AWS CDK and offers advanced features such as agent collaboration, vector database storage, observability via LangFuse, authentication, and a Streamlit web interface.

Check out the project here: GitHub Repo

Here is a video of the application in action:

Features

• 🤖 Multi-Agent Architecture: Built with AWS Bedrock Agents including a Supervisor Agent, RAG Agent and Tools Agent working together seamlessly as a multi-agent system
• 🔍 Knowledge Base Integration: Leverages AWS Bedrock Knowledge Base for intelligent processing of PDF documents via RAG
• 🔄 Tools Calling Integration: Custom Bedrock Action group for fetching and analyzing relevant clinical trials data from ClinicalTrials.gov using AWS Lambda
• 🧮 Code Interpreter: Built-in Code Interpreter capabilities for data analysis and visualization tasks that is able to generate images and HTML code
• 🎯 Prompt Engineering: Custom prompt templates and configurations for knowledge base responses and orchestration of the Bedrock Agents
• 📊 Vector Database Storage: Uses Amazon Aurora Vector Store for efficient storage and retrieval of embeddings for RAG
• 📈 Observability: Built-in monitoring and tracing with LangFuse integration for evaluation, tracing, and cost tracking
• 🛠️ Infrastructure as Code: Complete AWS CDK v2 for TypeScript setup with projen for consistent infrastructure management
• 🔐 Authentication: Integrated Cognito-based authentication system for secure access control to the application via username / password
• 💻 Streamlit App: Interactive web interface built with Streamlit for easy interaction with the agents deployed to AWS ECS Fargate
• 📚 OpenAPI Schema: Automatically generated from annotated code of your Lambda function using Lambda Powertools for Bedrock

Solution Design

Usage

• Login to your AWS account, go to AWS Bedrock Model catalog in eu-central-1 region and request access to the following models:
  • Titan Text Embeddings V2
  • Claude 3.5 Sonnet V1
• Eventually, you may need to increase your AWS account limits / service quota for Claude as described here
• Register a free LangFuse account for observability at LangFuse and create a new project called multi-agent-blueprint. Then create a new AWS Secret named langfuse/api that stores the public and private API key as JSON:

  {
    "LANGFUSE_PUBLIC_KEY": "<YOUR_PUBLIC_API_KEY>",
    "LANGFUSE_SECRET_KEY": "<YOUR_PRIVATE_API_KEY>"
  }

• Install uv and AWS CDK on your machine, then run the following commands:

  cd infra && npx projen && cd ../
  uv sync --frozen
  uv pip install -r src/app/requirements.dev.txt
  uv pip install -r src/tools/clinicaltrials/requirements.dev.txt

• Execute the following commands to deploy the infrastructure via AWS CDK:

  cd infra && cdk bootstrap
  DOCKER_DEFAULT_PLATFORM=linux/amd64 cdk deploy --require-approval never --all

• In the AWS Console under Bedrock Knowledge Bases, open the created knowledge base and make sure the data source sync is completed, otherwise trigger it manually
• Go to the AWS Console and create a new Cognito user with username and password. Use this to login to the Streamlit app over the HTTP URL that is printed out after the deployment is complete
• For local testing, set the Supervisor Agent ID and Alias ID as well as your LangFuse API keys in the .vscode/launch.json file, then run the APP launch configuration:

  {
    "env": {
      "SUPERVISOR_AGENT_ID": "<YOUR_SUPERVISOR_AGENT_ID>",
      "SUPERVISOR_AGENT_ALIAS_ID": "<YOUR_SUPERVISOR_AGENT_ALIAS_ID>",
      "LANGFUSE_SECRET_KEY": "<YOUR_LANGFUSE_SECRET_KEY>",
      "LANGFUSE_PUBLIC_KEY": "<YOUR_LANGFUSE_PUBLIC_KEY>"
    }
  }

• To regenerate the OpenAPI schema from the annotated code of your Lambda function, run the API SCHEMA launch configuration.

Possible Improvements

• Add end-to-end agent evaluation using a framework like AWS Agent Evaluation or LangFuse Evaluation Pipelines
• Add a reranker to the knowledge base to improve the quality of the RAG responses as shown in this code example
• Split the application into frontend and backend using FastAPI for the backend server, similar to this example that does the same for LangGraph
• Use a prompt router for Anthropic to route the user's query to the most appropriate agent, as explained in this example