Smart Expenses sample demonstration

Here is the architecture diagram of the application demonstrated in the Cloud Next 2021 talk "AI-powered applications with Google Cloud" (DEV202).

Learn how you can harness the power of serverless computing with AI services like Document AI. Build applications with speed, ease, and intelligence. See how AI-powered use cases, combined with Workflows and Cloud Functions, can help automate use cases at your enterprise. We’ll show you how to leverage pretrained models for form types like including receipts, invoice processing, and more. Developers will be able to increase productivity by spending less time worrying about infrastructure, and dedicate more time building applications for their businesses.

This demo application is a smart expense report application, consisting of two main screens:

a page for employees where they can select receipts and submit their expense reports,
a page for manager to review, then approve or reject submitted expense reports.

The project makes use of the following Google Cloud products:

Document AI to parse and analyse receipts to find meaningful information (supplier, line items, amounts, currency),
Workflows to organize the overall business process, invoking Document AI and data analysis function, to update status in Firestore, to create a callback waiting for the manager's approval,
Cloud Functions to make the link between frontend and backend, and to extract data from Document AI's JSON output,
Firestore to store all the data about the expense reports and receipts
Cloud Stroage to store pictures of receipts,
Firebase on the frontend to upload receipt pictures and to keep UI up-to-date with the realtime status of the expense report processing.

The project is using:

Vue.js as its frontend progressive JavaScript framework,
Shoelace as its library of web components,
Node.js for the code of the few cloud functions.

Some project setup notes

Deploying the sample within your own project

Certain URLs pointing at cloud storage buckets or function locations are hard-coded. To deploy the sample application in your own Google Cloud Project, you should update:

In workflow.yaml:

The bucket_input and bucket_output variables to point at globally unique GCS buckets.
In invoke_document_ai step, point at your location and at the right Document AI processor ID.
In process_annotations step, point at the URL of your cloud function analysing Document AI's JSON output.

In public/js/common.js:

Update the Firebase project configuration to point at your project (see Firebase documentation)

In public/js/admin.js:

Change the callbackFunctionUrl URL to point at the URL of your approval callback function.

In public/js/request.js:

Change the fnUrl URL to point at the URL of your workflow invocation function.

Deploying the workflow definition

gcloud workflows deploy batch-process-receipts \
    --location=europe-west4 \
    --source=workflow.yaml

Web assets with Firebase

firebase login

Deploy static assets to Firebase hosting

firebase deploy --only hosting

Run local server:

firebase serve

Functions deployment

Environment variables

export FUNCTION_REGION=europe-west1
export WORKFLOW_REGION=europe-west4
export WORKFLOW_NAME=batch-process-receipts

Function invoking the workflow from the web frontend:

gcloud functions deploy invoke-workflow \
  --region=${FUNCTION_REGION} \
  --source=./services/invoke-workflow \
  --runtime nodejs14 \
  --entry-point=invokeWorkflow \
  --set-env-vars PROJECT_ID=${GOOGLE_CLOUD_PROJECT},WORKFLOW_REGION=${WORKFLOW_REGION},WORKFLOW_NAME=${WORKFLOW_NAME} \
  --trigger-http \
  --allow-unauthenticated

Function calling the workflow callback from the web frontend:

gcloud functions deploy approval-callback \
  --region=${FUNCTION_REGION} \
  --source=./services/approval-callback \
  --runtime nodejs14 \
  --entry-point=approvalCallbackCall \
  --trigger-http \
  --allow-unauthenticated

Note: The service account used by the function calling the callback URL should have the Workflows Editor (or Workflows Admin) and Service Account Token Creator permissions.

Function analysing the Doc AI output

gcloud functions deploy process-annotations \
  --region=${FUNCTION_REGION} \
  --source=./services/process-annotations \
  --runtime nodejs14 \
  --entry-point=processAnnotations \
  --trigger-http

Firestore setup

To allow the web pages to access the data (in read-only mode) in Firestore, the security rules for Firestore should be updated with:

rules_version = '2';
service cloud.firestore {
  match /databases/{database}/documents {
    match /{document=**} {
      allow read, write: if false;
    }
  }
}

License

All solutions within this repository are provided under the Apache 2.0 license. Please see the LICENSE file for more detailed terms and conditions.

Disclaimer

This repository and its contents are not an official Google Product.

TonyChiangUSA/smart-expenses