AWS上のHugging Face smolagentsを用いたマルチモデルフレームワークによるエージェント型AI

*This post is cowritten by Jeff Boudier, Simon Pagezy, and Florent Gbelidji from Hugging Face.*

Agentic AI systems represent an evolution from conversational AI to autonomous agents capable of complex reasoning, tool usage, and code execution. Enterprise applications benefit from strategic deployment approaches tailored to specific needs. These needs include managed endpoints, which deliver auto-scaling capabilities, foundation model APIs to support complex reasoning, and containerized deployment options that support custom integration requirements.

Hugging Face smolagents is an open source Python library designed to make it straightforward to build and run agents using a few lines of code. We will show you how to build an agentic AI solution by integrating Hugging Face smolagents with Amazon Web Services (AWS) managed services. You’ll learn how to deploy a healthcare AI agent that demonstrates multi-model deployment options, vector-enhanced knowledge retrieval, and clinical decision support capabilities.

While we use healthcare as an example, this architecture applies to multiple industries where domain-specific intelligence and reliability are critical. The solution uses the model-agnostic, modality-agnostic, and tool-agnostic design of smolagents to orchestrate across Amazon SageMaker AI endpoints, Amazon Bedrock APIs, and containerized model servers.

Solution overview

Many AI systems face limitations with single-model approaches that can’t adapt to diverse enterprise needs. These systems often have rigid deployment options, inconsistent APIs across different AI services, and lack multi-model deployment options for optimal model selection.

This solution demonstrates how organizations can build AI systems that address these limitations. The solution allows deployment selection based on operational needs and provides consistent request and response formats across different AI backends and deployment methods. It generates contextual responses through medical knowledge integration and vector search, supporting deployment from development to production environments through containerized architecture.

This healthcare use case illustrates how the AI agent can process complex medical queries for six medications with clinical decision support and AWS security and compliance capabilities.

Architecture

The solution consists of the following services and features to deliver the agentic AI capabilities:

• SageMaker AI with BioM-ELECTRA-Large-SQuAD2 model for specialized medical queries and managed auto-scaling.

• Amazon Bedrock with Claude 3.5 Sonnet V2 by Anthropic for complex reasoning and foundation model access.

• Amazon OpenSearch Service for vector similarity matching and contextual knowledge retrieval with medical knowledge indexing.

• Amazon Elastic Container Service (Amazon ECS), with AWS Fargate, for serverless container orchestration and scalable deployment of the Python application that uses the smolagents library.

• AWS Identity and Access Management (IAM) for security and access control.

• Containerized model server with BioM-ELECTRA-Large-SQuAD2 for self-hosted model deployment.

The following diagram illustrates the solution architecture.

The architecture is a complete integration of the Hugging Face smolagents framework with AWS services. A client web interface connects to a healthcare agent container that orchestrates across three model backends: SageMaker AI with BioM-ELECTRA, Amazon Bedrock with Claude 3.5 Sonnet V2, and a containerized model server with BioM-ELECTRA. The solution includes a vector store powered by OpenSearch Service and a security layer with data encryption at rest and in transit. The security layer also handles IAM access control and authentication, and any medical disclaimers for regulatory compliance.

This solution supports deployment options through smolagents with each backend optimized for different scenarios:

• SageMaker AI for managed endpoints with auto-scaling and production workloads using Hugging Face Hub models.

• Amazon Bedrock for serverless access to foundation models and complex reasoning through AWS APIs.

• A containerized model server for self-hosted model deployment and tool integration from Hugging Face Hub.

The three backends implement Hugging Face Messages API compatibility, confirming consistent request and response formats regardless of the selected model service. Users select the appropriate backend based on their requirements—the solution provides deployment options rather than automatic routing.

The complete implementation is available in the sample-healthcare-agent-with-smolagents-on-aws GitHub repository.

Key benefits

The integration of Hugging Face smolagents with AWS managed services offers significant advantages for enterprise agentic AI deployments.

Deployment choice

Organizations can choose the optimal deployment for each use case: Amazon Bedrock for serverless access to foundation models and self-hosted containerized deployment for custom tool integration or SageMaker AI for specialized domain models. These options help to match specific workload requirements, rather than a one-size-fits-all approach.

Multi-model deployment options

Organizations can optimize their infrastructure choices without changing their agent logic. You can switch between containerized model server, SageMaker AI, and Amazon Bedrock without modifying your application code. This provides deployment options, while maintaining consistent agent behavior.

Code generation capabilities

The CodeAgent approach of smolagents streamlines multi-step operations through direct Python code generation and processing. The following comparison illustrates the multi-step operations of smolagents:

Multi-step JSON-based approach:

{
  "action": "search",
  "parameters": {"query": "drug interactions"},
  "next_action": {
    "action": "filter",
    "parameters": {"criteria": "severity > moderate"}
  }
}

smolagents CodeAgent:

# Search and filter in single code block
results = search_tool("drug interactions")
filtered_results = [r for r in results if r.severity > "moderate"]
final_answer(f"Found {len(filtered_results)} severe interactions: {filtered_results}")

The smolagents CodeAgent supports single code blocks to handle multi-step operations, reducing large language model (LLM) calls while streamlining agent development. It provides full control of agent logic across AWS service deployments.

Scalable architecture

By deploying the application on AWS, you gain access to security features and auto-scaling capabilities that help you meet organizational security requirements and maintain regulatory compliance. Running containerized workloads with Amazon ECS and Fargate helps you achieve reliable operations and optimize costs through automated resource scaling.

Let’s walk through implementing this solution.

Prerequisites

Before you deploy the solution, you need the following:

• An AWS account with appropriate permissions to create IAM roles, Amazon ECS clusters, and Amazon OpenSearch Service domains.

• AWS Command Line Interface (AWS CLI) version 2.0 or later installed and configured.

• Python 3.10 or later for running deployment scripts.

• Docker installed and running (required for production environments to provide secure code execution sandbox).

• Access to Amazon Bedrock, SageMaker AI, and OpenSearch Service in your AWS Region with appropriate IAM permissions to create and manage resources.

• For this implementation, we’re using Python 3.10+, smolagents framework, transformers 4.28.1+, PyTorch 2.0.0+, and boto3.

Run the following command to install the required Python packages:

pip install -r healthcare_ai_agent/phase_00_installation/requirements.txt

Configure environment variables

Set the required environment variables for your AWS Region and resource names before deploying the infrastructure.

• Set the following environment variables in your terminal:

export AWS_REGION=us-west-2
export SAGEMAKER_ENDPOINT_NAME=healthcare-qa-endpoint-1
export OPENSEARCH_DOMAIN=healthcare-vector-store
export OPENSEARCH_INDEX=medical-knowledge
export BEDROCK_MODEL_ID=anthropic.claude-3-5-sonnet-20241022-v2:0
export SAGEMAKER_MODEL_ID=sultan/BioM-ELECTRA-Large-SQuAD2
export CONTAINERIZED_MODEL_ID=sultan/BioM-ELECTRA-Large-SQuAD2

• Verify the variables are set:

echo $AWS_REGION
echo $SAGEMAKER_ENDPOINT_NAME

These environment variables are used throughout the deployment and testing processes. Verify that they’re set before proceeding to the next step.

Set up AWS infrastructure

Start by creating the foundational AWS infrastructure components using the SampleAWSInfrastructureManager class from the Smolagents_SageMaker_Bedrock_Opensearch.py implementation.

Deploy complete infrastructure (automated approach)

For automated deployment of the AWS infrastructure components, you can use the enhanced main function.

• Start the enhanced main function:

from Smolagents_SageMaker_Bedrock_Opensearch import enhanced_main
enhanced_main()
# Select option 1 for complete AWS infrastructure deployment

• The deployment automatically creates an Amazon ECS cluster, IAM roles, and an OpenSearch Service domain.

• Wait for deployment to complete (approximately 15–20 minutes).

Create individual AWS components (alternative approach)

If you prefer to create components individually, you can set up an OpenSearch Service domain for vector-enhanced knowledge retrieval and an Amazon ECS cluster for containerized deployment.

Both the OpenSearch Service domain and Amazon ECS cluster are automatically created as part of the complete AWS infrastructure deployment (Option 1 in enhanced_main). If you’ve already deployed the complete infrastructure, both components are ready and you can skip to the Deploy the Amazon SageMaker AI endpoint section.

Deploy the Amazon SageMaker AI endpoint

Deploy the BioM-ELECTRA-Large-SQuAD2 model to SageMaker AI for specialized medical query processing. An automated method (for deployment through the enhanced main) and a manual method (for deployment through the SageMaker AI endpoint) are provided.

To deploy using enhanced main (automated method)

from Smolagents_SageMaker_Bedrock_Opensearch import enhanced_main
# Start the enhanced main function
enhanced_main()
# Select option 2 for SageMaker endpoint deployment

To deploy the SageMaker AI endpoint (manual method)

• Verify your SageMaker environment variables are set (from the Configure environment variables section):

echo $SAGEMAKER_MODEL_ID
echo $SAGEMAKER_ENDPOINT_NAME

• Run the Amazon SageMaker deployment function:

from Smolagents_SageMaker_Bedrock_Opensearch import deploy_sagemaker_endpoint_safe
endpoint_name = deploy_sagemaker_endpoint_safe()

• The deployment uses the HuggingFaceModel with transformers 4.28.1, PyTorch 2.0.0, and ml.m5.xlarge instance type.

• Wait for the endpoint deployment to complete (approximately 5–10 minutes).

• Verify the endpoint deployment:

# Verify endpoint status
print(f"✅ Endpoint deployed: {endpoint_name}")

• The endpoint is configured for question-answering tasks with MAX_LENGTH=512 and TEMPERATURE=0.1.

Configure the multi-model backends

Configure the two additional backend options using the SampleTripleHealthcareAgent class for model selection based on operational needs.

Option 1 – Set up Amazon Bedrock access

Configure access to Amazon Bedrock for foundation model integration with Claude 3.5 Sonnet V2.

• Verify your Amazon Bedrock model configuration is set (from the Configure environment variables section):

echo $BEDROCK_MODEL_ID

• Access to Claude 3.5 Sonnet V2 is automatically available for your AWS account.

• You can verify model availability in the Amazon Bedrock console under Model catalog.

Initialize the medical knowledge base

Set up the medical knowledge database with six medications and vector embeddings in OpenSearch Service. You can use the enhanced_main() function, which provides an interactive menu for deployment tasks, or initialize manually using the SampleOpenSearchManager class.

To initialize using enhanced main (Automated method)

from Smolagents_SageMaker_Bedrock_Opensearch import enhanced_main
# Start the enhanced main function
enhanced_main()
# Select option 4 for OpenSearch initialization and medical knowledge indexing

To initialize the medical knowledge base (Manual method)

• Initialize the SampleOpenSearchManager:

from Smolagents_SageMaker_Bedrock_Opensearch import enhanced_main
# Start the enhanced main function
enhanced_main()
# Select option 4 for OpenSearch initialization and medical knowledge indexing

• The system creates the medical knowledge index with mappings for drug_name, content, and content_type fields.

• Medical knowledge for Metformin, Lisinopril, Atorvastatin, Amlodipine, Omeprazole, and Simvastatin is automatically indexed.

• Each medication includes side effects, monitoring requirements, and drug class information.

• The vector store supports similarity search with content type filtering for targeted queries.

• Verify the indexing completed successfully:

# Verify the indexing completed successfully 
print("Medical knowledge base initialized with 6 medications")

Option 2 – Deploy the containerized model server

After setting up the core infrastructure, deploy a containerized model server that provides self-hosted model deployment capabilities.

Deploy the containerized model server using the LocalContainerizedModelServer class with BioM-ELECTRA-Large-SQuAD2 for model deployment.

To deploy containers to Amazon ECS (automated method)

For automated deployment, use the containerized Amazon ECS deployment:

from Smolagents_SageMaker_Bedrock_Opensearch import enhanced_main
# Start the enhanced main function
enhanced_main()
# Select option 3 for ECS container deployment

To deploy the containerized model server (manual method)

• Verify your containerized model configuration is set (from the Configure environment variables step):

echo $CONTAINERIZED_MODEL_ID

• Initialize the containerized model server:

from Smolagents_SageMaker_Bedrock_Opensearch import LocalContainerizedModelServer
containerized_server = LocalContainerizedModelServer()
containerized_server.start_server()

The containerized model server uses Docker sandboxing for secure code execution.

• Verify that the containerized model server is running:

print(f"✅ Containerized Model Server status: {containerized_server.get_status()}")

The containerized model server includes