Vertex AI Agent Builder for a customer assistant

This project provides the demo of an Agent that simulates an insurance chatbot. The goal of this agent is the following :

"You are an assistant who can retrieve customer information, update the customer details and issue their membership card."

Disclaimer : This is not supported code by Google and is provided as is, without warranties of any kind.

This demonstration agent is built using Vertex AI Agent Builder from Google Cloud. Agent Apps are currently in Preview and shall not be used in production.

Architecture Design

This agent only demonstrates the integration of a simple OpenAPI tool and agent chaining, but you could also add Data stores or Functions when the agent is deployed.

Components

The content of this repo holds the following resources :

• Agent Application A Zip file that you can import in Vertex AI Agent Builder
• Web Service A web application using Python and Javascript that exposes a frontend website and an API.

Setup

Prerequisites

This application is deployed using an bash script that will deploy the content to a Google Repo and create a Cloud Run service using the gcloud official Google Cloud cli. Hence you need

• A Google Cloud project and access to the Google Cloud Console
• User rights on your project (the Cloud Run and Cloud Build API services will be enabled by the script)
• bash
• the gcloud utility

Set variables

In the setup.sh file, you will need to set the PROJECT_ID and LOCATION variables.

Deployment

The deployment will consist in two phases :

• 1/ Deploy a web service to Cloud Run for the frontend and API that simulates a CRM
• 2/ Import the Agent in the Agent Builder service

Deploy to Cloud Run

Download this repo and navigate to the appropriate folder.

Run the setup script :

user@penguin:~/med-agent$ ./setup.sh 

The setup script will display a Service URI such as : Service URL: https://customer-api-xxxxxxx-uc.a.run.app

You need to copy this URI and replace it in the script.js and openapi.yaml files so the API endpoint can be declared in the app. In the script.js file, you need to keep the /api/customers route.

Once you have replaced the URI in the files, run the setup script again using the same command :

user@penguin:~/med-agent$ ./setup.sh 

Deploy the Agent

• Go to the Google Cloud Console
• Navigate to the "Agent Builder" app (use the search bar if unsure)
• Go to the "Apps" section in the left menu and click the "Create App" button at the top
• Select the type "Agent"
• Give a "Name" to your agent and select a "Region". Ideally select the same region as where you deployed your Cloud Run service in the previous section.
• Once the App is created there is an icon at the top of your screen (two arrows up/down) that enable you to export/restore an app.
• Upload the agent contained in the med-app.zip file in this repo.

Using the models

Models vary as new ones are being released. This was tested using "gemini-1.0-pro-001".

Troubleshooting & known issues

It seems the Test Cases are currently meeting and issue so they may not be working upon deployment.