This repository contains the everything you need to deploy a smart and private question answering Chatbot from your Confluence data

Here is the demo of the app:

It uses the following architecture:

• 1. Prerequisites
• 2. Installation
  • 2.1. Bootstrap of the project
    • 2.1.1. Configure Terraform backend
    • 2.1.2. Bootstrap the project
  • 2.2. Push the docker image to the Artefact Registry
    • 2.2.1. Configure service account impersonation
    • 2.2.2. Add the secrets in GCP secret manager
  • 2.3. Deploy GCP ressources
    • 2.3.1. Deploy GCP Ressources
  • 2.4. Add data to your a Vertex Index
    • 2.4.1. Add environment variables for a local data intialization
    • 2.4.2. Deploy the Vertex Index Endpoint
• 3. Local development
     • 3.1. Prerequisites
     • 3.2. Install dependencies
     • 3.3. Run the app locally
     • 3.4. Understand the Chatbot logic with the notebooks
     • 3.5. Use docker and try the app on a container:
• 4. Delete ressources
     • 4.1. Delete the Vertex AI ressources
     • 4.2. Delete the terraform ressources
• 5. Troubleshooting
• 6. Roadmap
• 7. References

• Google Cloud CLI : Install Google Cloud SDK following the instructions here
• Terraform v>=1.5 : Follow the instructions here to install the Terraform CLI locally
• a GCP project with enabled billing

In the following commands, you will have to replace these variables with your own values:

• my-project-id
• my-tfstate-bucket : useful to share the tf.state file between multiple users
• my-user-email : the email you used to create your GCP account
• my-region : the region where you want to deploy the GCP ressources (ex: europe-west1)
• my-service-account : the name of the service account you want to create (ex: project-sa)
• my-registry-name : the name of the artificat registry that will be created (ex: rag-api)

| Prerequisite : GCP Admin Account

This step consistes of creating Terraform Service Account with the required roles (we used to deploy GCP ressources) and enable GCP APIs. If you are not owner of your project because you are using your company GCP account, you can ask your GCP admin to create a SA and give you the serviceAccountTokenCreator role.

To configure the Terraform backend first you need to:

• Create the GCS bucket to store the tf.state file The bucket name must be unique because the namespace is shared across all GCP projects.

gcloud init # Use your email and project settings
gcloud auth application-default login # Authenticate on GCP with admin account
gcloud storage buckets create gs://<my-tfstate-bucket> --project=<my-project-id> --location=<my-region> # You need storage.buckets.create permission (ex: roles/editor)

Update GCS bucket name & prefix :

cd terraform/bootstrap
vim backend.tf
# Edit <my-tfstate-bucket> with the name of the bucket created above
vim terraform.tfvars
# Edit the values of project_id, location, artifactregistry_name and service_account_name

terraform init # You need storage.buckets.list permission (ex: roles/storage.objectUser)
terraform apply -var gcp_account_email="<my-user-email>"

export GOOGLE_IMPERSONATE_SERVICE_ACCOUNT=<my-service-account>@<my-project-id>.iam.gserviceaccount.com

# Docker authentication
gcloud auth configure-docker <my-region>-docker.pkg.dev

# Build and push the image:
cd ../../ # Go back to the root directory
docker buildx build --platform linux/amd64 --push -t <my-region>-docker.pkg.dev/<my-project-id>/<my-registry-name>/gen-ai:latest .

vim terraform/terraform.tfvars
# Change the image name in the terraform/terraform.tfvars file with the one you created

To secure the Confluence identifiers, you need to add them in GCP secret manager. You need to store 4 secrets:

• CONFLUENCE_PRIVATE_API_KEY: the private API key of your Confluence account. Create a key here

• CONFLUENCE_SPACE_NAMES: the list of Confluence spaces names, useful for filters. You can find the space name in the URL of the space. For example, if the URL of your space is https://my-company.atlassian.net/wiki/spaces/SPACE_NAME/overview, the space name is SPACE_NAME.

• CONFLUENCE_URL

• CONFLUENCE_EMAIL_ADRESS

cd terraform
vim ./terraform.tfvars # Follow the instructions in the file to fill the variables
vim ./backend.tf # Update <my-tfstate-bucket> with the name of the bucket created above
terraform init
terraform apply # It will take approximatively 30 minutes to create the index

Now your Cloud Run service is deployed and you can access it with the URL displayed in the terraform output.

We have now to add data to the Vertex AI Search Index and deploy an endpoint to be able to query it.

Keep the two variables from the terraform output:

• my-index-id : the Vertex AI Search Index ID
• my-index-endpoint-id : the Vertex AI Search Index Endpoint ID They will be used in the next step.

The vector index is empty. We need to add data to it and deploy an endpoint to be able to query it. It will:

• collect urls from the Confluence Space you mentionned in the Confluence space names variable
• load each valid space url text and create chunks
• generate embeddings and save it to GCS

The data initialization in GCP compute engine will be available soon

cd ../src
cp ~/.config/gcloud/application_default_credentials.json ./credentials.json
cp ../.env.template .env
vim .env # add the required variables
# Install poetry and dependencies before ?
poetry run python -m upload_data.data_init
# You may have to wait up to an hour for the index to be updated.
# You can check the GCP Vector Search console and check the number of vectors in the index to see when it is updated.

Once the index is not empty and contains our data, we can deploy an endpoint to be able to query it.

gcloud ai index-endpoints deploy-index <my-index-endpoint-id> \
	--index=<my-index-id> \
	--deployed-index-id=ragdeployedindex \
	--display-name=ragendpoint \
	--region=<my-region>

You can go now go to the Cloud Run console and try the chatbot.

• Poetry

To work on a local environement, you first need to follow the preivous step or at least deploy a vector index and a deployed endpoint.

Create a virtual environment using poetry:

poetry install

You won't be able to use the secret manager in local, you have to copy the secrets in the .env file. If not already done

cd src
cp ~/.config/gcloud/application_default_credentials.json ./credentials.json
cp ../.env.template .env
vim .env # add the required variables

# To run the app in local:
poetry run python -m streamlit run src/chatbot/app.py

# At initialization, add new data:
poetry run python src/upload_data/data_init.py

# To update the index with new/modified/deleted Confluence data
poetry run python src/upload_data/data_update.py

make open-notebook

make demo
# Open http://localhost:8501/ in your browser

gcloud ai index-endpoints undeploy-index <my-index-endpoint-id> \
    --project=<my-project-id> \
    --deployed-index-id=ragdeployedindex \
    --region=europe-west1 \

terraform destroy

Cannot import Tensorflow

This can be caused in mac M1. Tensorflow requires to install tensorflow-metal with: poetry add tensorflow-metal.

Be careful when using make build-docker after this command because tensorflow-metal cannot be installed in the docker image.

To remove this dependency if needed in local: poetry remove tensorflow-metal

ERROR: (gcloud.ai.index-endpoints.deploy-index) FAILED_PRECONDITION: Deploying empty Index is not allowed.

This error can be caused by an empty index. You need to add data to the index before deploying an endpoint. If you have data in your storage, you have to update the index: make update-index. If the index vector number is not 0 in the GCP console, try creating another index endpoint and deploy it with this new endpoint.

• Switch to Hexagonale Architecture
• Design a Cloud Build CI/CD
• Initialize data on GCP Compute Engine
• Add the Google Drive loader with LangChain or our custom implementation
• Improve Retriever with our custom implementation
• Integrate the evaluation process directly in GCP
• Export the dashboard from LangSmith to monitor RAG process in GCP

Article: Understand the RAG process with an example

Code: GCP notebooks to understand the ingestion process and Vector Search usage

Doc: GCP Vector Search documentation on ingestion data format

Doc: Understand the Vector Search filtering process

Doc: Connect a PubSub subscription to BigQuery with a schema to bypass Dataflow

Vidéo: How to implement IAP in your Cloud Run service