Skip to content

AWS AI Stack – A ready-to-use, full-stack boilerplate project for building serverless AI applications on AWS

Notifications You must be signed in to change notification settings

serverless/aws-ai-stack

Repository files navigation

AWS AI Stack

AWS AI Stack – A ready-to-use, full-stack boilerplate project for building serverless AI applications on AWS. A great fit for those seeking a trusted AWS foundation for AI apps and access to powerful LLM models via Bedrock ​​that keep your app’s data separate from model providers.

View the Live Demo – awsaistack.com

Use this as a boilerplate project to create an AI Chat bot, authentication services, business logic, async workers, all on AWS Lambda, API Gateway, DynamoDB, and EventBridge.

This is a true serverless architecture, so you only pay for what you use, not for idle time. Some services, like DynamoDB, or AWS Bedrock trained models, may have additional storage costs.

serverless-framework-v4-aws-ai-stack-screenshots

Features

  • Full-Stack Application
    • Backend: API (AWS API Gateway V2, AWS Lambda), Event-driven architecture (AWS Event-Bridge, AWS Lambda), Database (AWS DynamoDB), AI (AWS Bedrock)
    • Frontend: Vanilla React app.
  • AI Chat & Streaming Responses
    • Full serverless AI Chat architecture w/ streaming responses on AWS Lambda.
  • Multiple AI Models & Data Privacy
    • Use one or multiple models via AWS Bedrock: Claude 3.5 Sonnet, Llama3.1, Mistral Large 2, and many more.
    • App data never leaves AWS and is not sent to model providers.
  • 100% Serverless
    • This is a true serverless architecture. It auto-scales and you only pay when users use it. Some services may have additional storage costs.
  • Custom Domain Names
    • Custom domain names for API Gateway services using the serverless-domain-manager plugin
    • Custom domain names for Lambda services using CloudFront Distributions
  • API & Event-Driven
    • Express.js API placeholder service for your business logic
    • Shared EventBridge to public & subscribe to events
    • Worker service to process events from EventBridge
  • Built-In Authentication
    • API Gateway authorizer
    • Login & Registration API on Lambda with Express.js
    • DynamoDB table to store user information
    • Shared library to provide JWT token authentication
    • Frontend website that uses login & registration API
  • Multi-Environment
    • Shared configuration for all services.
    • Separated configuration for different environments.
  • Domain Oriented Architecture
    • This project is domain-oriented so you can easily remove the pieces you don't need, like AI Chat, authentication, etc.
  • CI/CD with Github Action
    • Github Actions to deploy the services to prod.
    • Github Actions to deploy PRs & remove services after merge.

Getting Started

1. Install dependencies

Install Serverless Framework

npm i -g serverless

Install NPM dependencies

This project is structured as a monorepo with multiple services. Each service has its own package.json file, so you must install the dependencies for each service. Running npm install in the root directory will install the dependencies for all services.

npm install

Setup AWS Credentials

If you haven't already, setup your AWS Credentials. You can follow the AWS Credentials doc for step-by-step instructions.

2. Enable AWS Bedrock Models

This example requires the meta.llama3-70b-instruct-v1:0 AWS Bedrock Model to be enabled. By default, AWS does not enable these models, you must go to the AWS Console and individually request access to the AI Models.

There is no cost to enable the models, but you must request access to use them.

Upon request, it may take a few minutes for AWS to enable the model. Once they are enabled, you will receive an email from AWS confirming the model is enabled.

Some users have reported issues with getting models enabled on AWS Bedrock. Make sure you have sufficient permissions in AWS to enable the models first. Often, AWS accounts that are new or have not historically had a monthly invoice over a few dollars may require contacting AWS to enable models.

3. Deploy & start developing

Now you are ready to deploy the services. This will deploy all the services to your AWS account. You can deploy the services to the default stage, which is the default stage for development.

Deploy the services

serverless deploy

At this point the service is live. When running the serverless deploy command, you will see the output of the services that were deployed. One of those services is the web service, which is the website service. To view the app, go to the URL in the endpoint: ANY - section for the web service.

Deploying "web" to stage "dev" (us-east-1)

endpoint: ANY - https://ps5s7dd634.execute-api.us-east-1.amazonaws.com
functions:
  app: web-dev-app (991 kB)

Once you start developing it is easier to run the service locally for faster iteration. We recommend using Serverless Dev Mode. You can run Dev Mode for individual services. This emulates Lambda locally and proxies requests to the real service.

serverless auth dev

Once done, you can redeploy individual services using the serverless command with the service name.

serverless auth deploy

The website service is a static website that is served from an AWS Lambda function. As such, it can run locally without needing to use Dev Mode. However, it has a dependency on the AI Chat service and the Auth service, so you must configure environment variables locally.

# If you have the jq CLI command installed you can use that with the --json flag
# on serverless info to get the URLs from the deployed services. If you do not
# have jq installed, you can get the URLs by running "serverless auth info" and
# "serverless ai-chat info" and copying the URLs manually into the environment
# variables.
export VITE_CHAT_API_URL=$(serverless aiChatApi info --json | jq -r '.outputs[] | select(.OutputKey == "ChatApiUrl") | .OutputValue')
export VITE_AUTH_API_URL=$(serverless auth info --json | jq -r '.outputs[] | select(.OutputKey == "AuthApiUrl") | .OutputValue')

# now you can run the local development server
cd website/app
npm run build

4. Prepare & release to prod

Now that the app is up and running in a development environment, lets get it ready for production by setting up a custom domain name, and setting a new shared secret for JWT token authentication.

Setup Custom Domain Name (optional)

This project is configured to use custom domain names. For non prod deployments this is disabled. Deployments to prod are designed to use a custom domain name and require additional setup:

Register the domain name & create a Route53 hosted zone

If you haven't already, register a domain name, and create a Route53 hosted zone for the domain name.

https://us-east-1.console.aws.amazon.com/route53/v2/hostedzones?region=us-east-1#

Create a Certificate in AWS Certificate Manager

A Certificate is required in order to use SSL (https) with a custom domain name. AWS Certificate Manager (ACM) provides free SSL certificates for use with your custom domain name. A certificate must first be requested, which requires verification, and may take a few minutes.

https://us-east-1.console.aws.amazon.com/acm/home?region=us-east-1#/certificates/list

After you have created the certificate, you must validate the certificate by following the instructions in the AWS Console. This may require adding a CNAME record to your DNS provider.

This example uses a Certificate with the following full qualified domain names:


awsaistack.com
\*.awsaistack.com

The base domain name, awsaistack.com is used for the website service to host the static website. The wildcard domain name, *.awsaistack.com is used for the API services, api.awsaistack.com, and chat.awsaistack.com.

Update serverless-compose.yml

  • Update the stages.prod.params.customDomainName to your custom domain name.
  • Update the stages.prod.params.customDomainCertificateARN to the ARN of the certificate you created in ACM.

Create the secret for JWT token authentication

Authentication is implemented using JWT tokens. A shared secret is used to sign the JWT tokens when a user logs in. The secret is also used to verify the JWT tokens when a user makes a request to the API. It is important that this secret is kept secure and not shared.

In the serverless-compose.yml file, you'll see that the sharedTokenSecret is set to "DEFAULT" in the stages.default.params section. This is a placeholder value that is used when the secret is not provided in non-prod environments.

The prod stage uses the ${ssm} parameter to retrieve the secret from AWS Systems Manager Parameter Store.

Generate a random secret and store it in the AWS Systems Manager Parameter Store with a key like /serverless-ai-service/shared-token, and set it in the stages.prod.params.sharedTokenSecret parameter in the serverless-compose.yml file:

sharedTokenSecret: ${ssm:/serverless-ai-service/shared-token}

Deploy to prod

Once you've setup the custom domain name (optional), and created the secret, you are ready to deploy the service to prod.


serverless deploy --stage prod

Now you can use the service by visiting your domain name, or https://awsaistack.com. This uses the Auth service to login and register users, the AI Chat service to interact with the AI Chat bot.

Architectural Overview

Serverless & Costs

This example uses serverless services like AWS Lambda, API Gateway, DynamoDB, EventBridge, and CloudFront. These services are designed to scale with usage, and you only pay for what you use. This means you do not pay for idle, and only pay for the resources you consume. If you have 0 usage, you will have $0 cost.

If you are using the custom domain names, it will require Route53 which has a fixed monthly cost.

Compose

This example uses Serverless Compose to share configuration across all services.

It defines the global parameters in the serverless-compose.yml file under stages.default.params and stages.prod.params. These parameters are used across all services to provide shared configuration.

It also uses CloudFormation from services to set parameters on other services. For example, the auth service publishes the CloudFormation Output AuthApiUrl, which is used by the website service.

web:
  path: ./website
  params:
    authApiUrl: ${auth.AuthApiUrl}

Using Serverless Compose also allows you to deploy all services with a single command, serverless deploy.

Authentication SDK Library

The auth service contains a shared client library that is used by the other services to validate the JWT token. This library is defined as an NPM package and is used by the ai-chat-api and business-api services and included using relative paths in the package.json file.

Authentication (api.awsaistack.com/auth)

The auth service is an Express.js-based API service that provides login and registration endpoints. It uses a DynamoDB table to store user information and uses JWT tokens for authentication.

Upon login or registration, the service returns a JWT token. These APIs are used by the website service to authenticate users. The token is stored in localstorage and is used to authenticate requests to the ai-chat-api and business-api services.

The ai-chat-api service uses AWS Lambda Function URLs instead of API Gateway, in order to support streaming responses. As such, it uses the Auth class from auth-sdk to validate the JWT token, instead of using an API Gateway authorizer.

The auth service also publishes the CloudFormation Output AuthApiUrl, which is used by the website service to make requests to the auth service.

AI Chat (chat.awsaistack.com)

In most cases APIs on AWS Lambda use the API Gateway to expose the API. However, the ai-chat-api service uses Lambda Function URLs instead of API Gateway, in order to support streaming responses as streaming responses are not supported by API Gateway.

Since the ai-chat-api service does not use API Gateway, it does not support custom domain names natively. Instead, it uses a CloudFront Distribution to support a custom domain name.

To provide the AI Chat functionality, the service uses the AWS Bedrock Models service to interact with the AI Chat bot. The requests from the frontend (via the API) are sent to the AWS Bedrock Models service, and the streaming response from Bedrock is sent back to the frontend via the streaming response.

The AWS Bedrock AI Model is selected using the modelId parameter in the ai-chat-api/serverless.yml file.

stages:
  default:
    params:
      modelId: meta.llama3-70b-instruct-v1:0

The AI Chat service also implements a simple throttling schema to limit cost exposure when using AWS Bedrock. It implements a monthly limit for the number of requests per user and a global monthly limit for all users. It uses a DynamoDB Table to persist the request counts and other AI usage metrics.

The inline comments provider more details on this mechanism as well as ways to customize it to use other metrics, like token usage.

stages:
  default:
    params:
      throttleMonthlyLimitUser: 10
      throttleMonthlyLimitGlobal: 100

Website (awsaistack.com)

The website service is a simple Lambda function which uses Express to serve static assets. The service uses the serverless-plugin-scripts plugin to run the npm run build command to build the website before deploying.

The build command uses the parameters to set the REACT_APP_* environment variables, which are used in the React app to configure the API URLs.

The frontend website is built using React. It uses the auth service to login and register uses, and uses the ai-chat-api to interact with the AI Chat bot API.

Business API (api.awsaistack.com/business)

This is an Express.js-based API service that provides a placeholder for your business logic. It is configured to use the same custom domain name as the auth service, but with a different base path (/business).

The endpoints are protected using the express-jwt middleware, which uses the JWT token provided by the auth service to authenticate the user.

Business Worker

This is a placeholder function for your business logic for processing asynchronous events. It subscribes to events on the EventBridge and processes the events.

Currently this subscribes to the auth.register event, which is published by the auth service when a user registers.

Both the Business Worker and the Auth service therefore depend on the EventBridge which is provisioned in the event-bus service.

Custom Domain Name

The services which use API Gateway use the serverless-domain-manager plugin to setup the custom domain name. More details about the plugin can be found on the serverless-domain-manager plugin page.

The api-ai-chat service uses Lambda Function URLs instead of API Gateway, so custom domain name is supported by creating a CloudFront Distribution with the custom domain name and the Lambda Function URL as the origin.

The business-api and auth APIs both use the same custom domain name. Instead of sharing an API Gateway, they are configured to use the same domain name with different base paths, one for each service.

API Usage

Below are a few simple API requests using the curl command.

User Registration API

curl -X POST https://api.awsaistack.com/auth/register \
  -H 'Content-Type: application/json' \
  -d '{"email": "me@example.com", "password": "password"}'

User Login API

curl -X POST https://api.awsaistack.com/auth/login \
  -H 'Content-Type: application/json' \
  -d '{"email": "me@example.com", "password": "password"}'

If you have jq installed, you can wrap the login request in a command to set the token as an environment variable so you can use the token in subsequent requests.

export SERVERLESS_EXAMPLE_TOKEN=$(curl -X POST https://api.awsaistack.com/auth/login \
  -H 'Content-Type: application/json' \
  -d '{"email": "me@example.com", "password": "password"}' \
| jq -r '.token')

Chat API

You can also use the Chat API directly; however, the response payload is a a stream of JSON objects containing the response and other metadata. Each buffer may also contain multiple JSON objects.

This endpoint is authenticated and requires the JWT token from the login API.

curl -N -X POST https://chat.awsaistack.com/ \
  -H 'Content-Type: application/json' \
  -H "Authorization: Bearer $SERVERLESS_EXAMPLE_TOKEN" \
  -d '[{"role":"user","content":[{"text":"What makes the serverless framework so great?"}]}]'

Business Logic API

This endpoint is also authenticated and requires the JWT token from the login API. The response is a simple message.

curl -X GET https://api.awsaistack.com/business/ \
  -H 'Content-Type: application/json' \
  -H "Authorization: Bearer $SERVERLESS_EXAMPLE_TOKEN"

Alternative Design Considerations

Sharing Domain Name between CloudFront and API Gateway

The Chat API uses CloudFront Distributions to add support for custom domain names to the AWS Lambda Function URL, as it is not natively supported. The Auth & Business APIs on the other hand use API Gateway which supports custom domain names natively. However, an API Gateway and a CloudFront Distribution do not support using the same hostname as they both require a CNAME record.

For these two services to share the same domain name, consider using the CloudFront distribution to proxy the API Gateway requests. This would allow both services to use the same domain name, and would also allow the Chat API to use the same domain name as the other services.

Using API Gateway to share a custom domain name

In this configuration, the Auth and Business APIs use the paths /auth and /business respectively on api.awsaistack.com. The Custom Domain Name Path Mapping was used in the Custom Domain Name support in API Gateway to use the same domain name but shared across multiple API Gateway instances.

Alternatively, you can use a single API Gateway and map the paths to the respective services. This would allow you to use the same domain name for multiple services, and would also allow you to use the same authorizer for all the services. However, sharing an API Gateway instance may have performance implications at scale, which is why this example uses separate API Gateway instances for each service.

Schema validation

The auth, business-api, and chat-api all validate the user input, and in the case of chat-api use Zod to validate the schema. Consider including schema validation on all API requests using a library like Zod, and/or Express.js middleware.

Static website hosting

This example, for simplicity, hosts the static assets from an AWS Lambda Function. This is not recommended for production, and you should consider using a static website hosting service like S3 or CloudFront to host your website. Consider using one of the following plugins to deploy your website:

Using Lambda Authorizers

This example uses a custom authorization method using JWT tokens for the ai-chat-api service, which doesn't use API Gateway.

The business-api is based on Express.js and uses the authMiddleware method in the auth-sdk to validate the JWT token.

API Gateway supports Lambda Authorizers which can be used to validate JWT tokens before the request is passed to the Lambda Function. This is a more robust solution than the custom method used in this example, and should be considered for production services. This method will not work for the ai-chat-api service as it does not use API Gateway.

Deploying with CI/CD

Using Github Actions this example deploys all the services using Serverless Compose. This ensures that any changes to the individual services or the serverless-compose.yml will reevaluate the interdependent parameters. However, all services are redeployed on any change in the repo, which may not be necessary.

Consider using a more fine-grained approach to deploying services, such as only deploying the services that have changed by using the serverless <service> deploy command.

About

AWS AI Stack – A ready-to-use, full-stack boilerplate project for building serverless AI applications on AWS

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages