This tutorial shows you how to integrate a large language model (LLM) application based on retrieval-augmented generation (RAG) with PDF files that you upload to a Cloud Storage bucket.
This guide uses a database as a storage and semantic search engine that holds the representations (embeddings) of the uploaded documents. You use the Langchain framework to interact with the embeddings and you use Gemini models available through Vertex AI.
Langchain is a popular open-source Python framework that simplifies many machine learning tasks and has interfaces to integrate with different vector databases and AI services.
This tutorial is intended for cloud platform administrators and architects, ML engineers, and MLOps (DevOps) professionals interested in deploying RAG LLM applications to GKE and Cloud Storage.
Objectives
In this tutorial, you learn how to:
- Build and deploy an application to create and store document embeddings in a vector database.
- Automate the application to trigger new document uploads to a Cloud Storage bucket.
- Deploy a chatbot application that uses semantic search to answer questions based on the document content.
Deployment architecture
In this tutorial, you create a Cloud Storage bucket, an Eventarc trigger, and the following Services:
embed-docs
: Eventarc triggers this Service every time a user uploads a new document to the Cloud Storage bucket. The Service starts a Kubernetes Job which creates embeddings for the uploaded document and inserts the embeddings into a vector database.chatbot
: This Service answers natural language questions about the uploaded documents using semantic search and the Gemini API.
The following diagram shows the process of uploading and vectorizing documents:
In the diagram, the user uploads files into the Cloud Storage bucket.
Eventarc subscribes to object metadataUpdated
events for the
bucket and uses Eventarc's event forwarder, which is a Kubernetes
workload, to call the embed-docs
Service when you upload a new document. The
Service then creates embeddings for the uploaded document. The embed-docs
Service stores the embeddings in a vector database using the Vertex AI
embedding model.
The following diagram shows the process of asking questions about the uploaded
document content using the chatbot
Service:
Users can ask questions using natural language, and the chatbot generates answers based solely on the content of the uploaded files. The chatbot retrieves context from the vector database using semantic search, then sends the question and context to Gemini.
Costs
In this document, you use the following billable components of Google Cloud:
To generate a cost estimate based on your projected usage,
use the pricing calculator.
When you finish the tasks that are described in this document, you can avoid continued billing by deleting the resources that you created. For more information, see Clean up.
Before you begin
In this tutorial, you use Cloud Shell to run commands. Cloud Shell is a shell environment for managing resources hosted on Google Cloud. Cloud Shell comes preinstalled with the Google Cloud CLI, kubectl, and Terraform command-line tools. If you don't use Cloud Shell, install the Google Cloud CLI.
- Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.
- Install the Google Cloud CLI.
-
To initialize the gcloud CLI, run the following command:
gcloud init
-
Create or select a Google Cloud project.
-
Create a Google Cloud project:
gcloud projects create PROJECT_ID
Replace
PROJECT_ID
with a name for the Google Cloud project you are creating. -
Select the Google Cloud project that you created:
gcloud config set project PROJECT_ID
Replace
PROJECT_ID
with your Google Cloud project name.
-
-
Make sure that billing is enabled for your Google Cloud project.
-
Enable the Vertex AI, Cloud Build, Eventarc, Artifact Registry APIs:
gcloud services enable aiplatform.googleapis.com
cloudbuild.googleapis.com eventarc.googleapis.com artifactregistry.googleapis.com - Install the Google Cloud CLI.
-
To initialize the gcloud CLI, run the following command:
gcloud init
-
Create or select a Google Cloud project.
-
Create a Google Cloud project:
gcloud projects create PROJECT_ID
Replace
PROJECT_ID
with a name for the Google Cloud project you are creating. -
Select the Google Cloud project that you created:
gcloud config set project PROJECT_ID
Replace
PROJECT_ID
with your Google Cloud project name.
-
-
Make sure that billing is enabled for your Google Cloud project.
-
Enable the Vertex AI, Cloud Build, Eventarc, Artifact Registry APIs:
gcloud services enable aiplatform.googleapis.com
cloudbuild.googleapis.com eventarc.googleapis.com artifactregistry.googleapis.com -
Grant roles to your user account. Run the following command once for each of the following IAM roles:
eventarc.admin
gcloud projects add-iam-policy-binding PROJECT_ID --member="USER_IDENTIFIER" --role=ROLE
- Replace
PROJECT_ID
with your project ID. -
Replace
USER_IDENTIFIER
with the identifier for your user account. For example,user:myemail@example.com
. - Replace
ROLE
with each individual role.
- Replace
Create a cluster
Create a Qdrant, Elasticsearch, or Postgres cluster:
Qdrant
Follow the instructions in Deploy a Qdrant vector database on GKE to create a Qdrant cluster running on an Autopilot mode or Standard mode GKE cluster.
Elasticsearch
Follow the instructions in Deploy an Elasticsearch vector database on GKE to create an Elasticsearch cluster running on an Autopilot mode or Standard mode GKE cluster.
PGVector
Follow the instructions in Deploy a PostgreSQL vector database on GKE to create a Postgres cluster with PGVector running on an Autopilot mode or Standard mode GKE cluster.
Weaviate
Follow the instructions to Deploy a Weaviate vector database on GKE to create a Weaviate cluster running on an Autopilot or Standard mode GKE cluster.
Set up your environment
Set up your environment with Cloud Shell:
Set environment variables for your project:
Qdrant
export PROJECT_ID=PROJECT_ID export KUBERNETES_CLUSTER_PREFIX=qdrant export REGION=us-central1 export DB_NAMESPACE=qdrant
Replace
PROJECT_ID
with your Google Cloud project ID.Elasticsearch
export PROJECT_ID=PROJECT_ID export KUBERNETES_CLUSTER_PREFIX=elasticsearch export REGION=us-central1 export DB_NAMESPACE=elastic
Replace
PROJECT_ID
with your Google Cloud project ID.PGVector
export PROJECT_ID=PROJECT_ID export KUBERNETES_CLUSTER_PREFIX=postgres export REGION=us-central1 export DB_NAMESPACE=pg-ns
Replace
PROJECT_ID
with your Google Cloud project ID.Weaviate
export PROJECT_ID=PROJECT_ID export KUBERNETES_CLUSTER_PREFIX=weaviate export REGION=us-central1 export DB_NAMESPACE=weaviate
Replace
PROJECT_ID
with your Google Cloud project ID.Verify that your GKE cluster is running:
gcloud container clusters list --project=${PROJECT_ID} --region=${REGION}
The output is similar to the following:
NAME LOCATION MASTER_VERSION MASTER_IP MACHINE_TYPE NODE_VERSION NUM_NODES STATUS [KUBERNETES_CLUSTER_PREFIX]-cluster us-central1 1.30.1-gke.1329003 <EXTERNAL IP> e2-standard-2 1.30.1-gke.1329003 6 RUNNING
Clone the sample code repository from GitHub:
git clone https://github.com/GoogleCloudPlatform/kubernetes-engine-samples
Navigate to the
databases
directory:cd kubernetes-engine-samples/databases
Prepare your infrastructure
Create an Artifact Registry repository, build Docker images, and push Docker images to Artifact Registry:
Create an Artifact Registry repository:
gcloud artifacts repositories create ${KUBERNETES_CLUSTER_PREFIX}-images \ --repository-format=docker \ --location=${REGION} \ --description="Vector database images repository" \ --async
Set the
storage.objectAdmin
andartifactregistry.admin
permissions on the Compute Engine service account to use Cloud Build to build and push Docker images for theembed-docs
andchatbot
Services.export PROJECT_NUMBER=PROJECT_NUMBER gcloud projects add-iam-policy-binding ${PROJECT_ID} \ --member="serviceAccount:${PROJECT_NUMBER}-compute@developer.gserviceaccount.com" \ --role="roles/storage.objectAdmin" gcloud projects add-iam-policy-binding ${PROJECT_ID} \ --member="serviceAccount:${PROJECT_NUMBER}-compute@developer.gserviceaccount.com" \ --role="roles/artifactregistry.admin"
Replace
PROJECT_NUMBER
with your Google Cloud project number.Build Docker images for the
embed-docs
andchatbot
Services. Theembed-docs
image contains Python code for both the application that receives Eventarc forwarder requests and the embedding job.Qdrant
export DOCKER_REPO="${REGION}-docker.pkg.dev/${PROJECT_ID}/${KUBERNETES_CLUSTER_PREFIX}-images" gcloud builds submit qdrant/docker/chatbot --region=${REGION} \ --tag ${DOCKER_REPO}/chatbot:1.0 --async gcloud builds submit qdrant/docker/embed-docs --region=${REGION} \ --tag ${DOCKER_REPO}/embed-docs:1.0 --async
Elasticsearch
export DOCKER_REPO="${REGION}-docker.pkg.dev/${PROJECT_ID}/${KUBERNETES_CLUSTER_PREFIX}-images" gcloud builds submit elasticsearch/docker/chatbot --region=${REGION} \ --tag ${DOCKER_REPO}/chatbot:1.0 --async gcloud builds submit elasticsearch/docker/embed-docs --region=${REGION} \ --tag ${DOCKER_REPO}/embed-docs:1.0 --async
PGVector
export DOCKER_REPO="${REGION}-docker.pkg.dev/${PROJECT_ID}/${KUBERNETES_CLUSTER_PREFIX}-images" gcloud builds submit postgres-pgvector/docker/chatbot --region=${REGION} \ --tag ${DOCKER_REPO}/chatbot:1.0 --async gcloud builds submit postgres-pgvector/docker/embed-docs --region=${REGION} \ --tag ${DOCKER_REPO}/embed-docs:1.0 --async
Weaviate
export DOCKER_REPO="${REGION}-docker.pkg.dev/${PROJECT_ID}/${KUBERNETES_CLUSTER_PREFIX}-images" gcloud builds submit weaviate/docker/chatbot --region=${REGION} \ --tag ${DOCKER_REPO}/chatbot:1.0 --async gcloud builds submit weaviate/docker/embed-docs --region=${REGION} \ --tag ${DOCKER_REPO}/embed-docs:1.0 --async
Verify the images:
gcloud artifacts docker images list $DOCKER_REPO \ --project=$PROJECT_ID \ --format="value(IMAGE)"
The output is similar to the following:
$REGION-docker.pkg.dev/$PROJECT_ID/${KUBERNETES_CLUSTER_PREFIX}-images/chatbot $REGION-docker.pkg.dev/$PROJECT_ID/${KUBERNETES_CLUSTER_PREFIX}-images/embed-docs
Deploy a Kubernetes Service Account with permissions to run Kubernetes Jobs:
Qdrant
sed "s/<PROJECT_ID>/$PROJECT_ID/;s/<CLUSTER_PREFIX>/$KUBERNETES_CLUSTER_PREFIX/" qdrant/manifests/05-rag/service-account.yaml | kubectl -n qdrant apply -f -
Elasticsearch
sed "s/<PROJECT_ID>/$PROJECT_ID/;s/<CLUSTER_PREFIX>/$KUBERNETES_CLUSTER_PREFIX/" elasticsearch/manifests/05-rag/service-account.yaml | kubectl -n elastic apply -f -
PGVector
sed "s/<PROJECT_ID>/$PROJECT_ID/;s/<CLUSTER_PREFIX>/$KUBERNETES_CLUSTER_PREFIX/" postgres-pgvector/manifests/03-rag/service-account.yaml | kubectl -n pg-ns apply -f -
Weaviate
sed "s/<PROJECT_ID>/$PROJECT_ID/;s/<CLUSTER_PREFIX>/$KUBERNETES_CLUSTER_PREFIX/" weaviate/manifests/04-rag/service-account.yaml | kubectl -n weaviate apply -f -
When using Terraform to create the GKE cluster and have
create_service_account
set as true, a separate service account will be created and used by the cluster and nodes. Grantartifactregistry.serviceAgent
role to this Compute Engine service account to allow the nodes to pull image from the Artifact Registry created forembed-docs
andchatbot
.export CLUSTER_SERVICE_ACCOUNT=$(gcloud container clusters describe ${KUBERNETES_CLUSTER_PREFIX}-cluster \ --region=${REGION} \ --format="value(nodeConfig.serviceAccount)") gcloud projects add-iam-policy-binding ${PROJECT_ID} \ --member="serviceAccount:${CLUSTER_SERVICE_ACCOUNT}" \ --role="roles/artifactregistry.serviceAgent"
Without granting access to the service account, your nodes might experience permission issue when trying to pull image from the Artifact Registry when deploying the
embed-docs
andchatbot
Services.Deploy a Kubernetes Deployment for the
embed-docs
andchatbot
Services:Qdrant
sed "s|<DOCKER_REPO>|$DOCKER_REPO|" qdrant/manifests/05-rag/chatbot.yaml | kubectl -n qdrant apply -f - sed "s|<DOCKER_REPO>|$DOCKER_REPO|" qdrant/manifests/05-rag/docs-embedder.yaml | kubectl -n qdrant apply -f -
Elasticsearch
sed "s|<DOCKER_REPO>|$DOCKER_REPO|" elasticsearch/manifests/05-rag/chatbot.yaml | kubectl -n elastic apply -f - sed "s|<DOCKER_REPO>|$DOCKER_REPO|" elasticsearch/manifests/05-rag/docs-embedder.yaml | kubectl -n elastic apply -f -
PGVector
sed "s|<DOCKER_REPO>|$DOCKER_REPO|" postgres-pgvector/manifests/03-rag/chatbot.yaml | kubectl -n pg-ns apply -f - sed "s|<DOCKER_REPO>|$DOCKER_REPO|" postgres-pgvector/manifests/03-rag/docs-embedder.yaml | kubectl -n pg-ns apply -f -
Weaviate
sed "s|<DOCKER_REPO>|$DOCKER_REPO|" weaviate/manifests/04-rag/chatbot.yaml | kubectl -n weaviate apply -f - sed "s|<DOCKER_REPO>|$DOCKER_REPO|" weaviate/manifests/04-rag/docs-embedder.yaml | kubectl -n weaviate apply -f -
Enable Eventarc triggers for GKE:
gcloud eventarc gke-destinations init
When prompted, enter
y
.Deploy the Cloud Storage bucket and create an Eventarc trigger using Terraform:
export GOOGLE_OAUTH_ACCESS_TOKEN=$(gcloud auth print-access-token) terraform -chdir=vector-database/terraform/cloud-storage init terraform -chdir=vector-database/terraform/cloud-storage apply \ -var project_id=${PROJECT_ID} \ -var region=${REGION} \ -var cluster_prefix=${KUBERNETES_CLUSTER_PREFIX} \ -var db_namespace=${DB_NAMESPACE}
When prompted, type
yes
. It might take several minutes for the command to complete.Terraform creates the following resources:
- A Cloud Storage bucket to upload the documents
- An Eventarc trigger
- A Google Cloud Service Account named
service_account_eventarc_name
with permission to use Eventarc. - A Google Cloud Service Account named
service_account_bucket_name
with permission to read the bucket and access Vertex AI models.
The output is similar to the following:
... # Several lines of output omitted Apply complete! Resources: 15 added, 0 changed, 0 destroyed. ... # Several lines of output omitted
Load documents and run chatbot queries
Upload the demo documents and run queries to search over the demo documents using the chatbot:
Upload the example
carbon-free-energy.pdf
document to your bucket:gsutil cp vector-database/documents/carbon-free-energy.pdf gs://${PROJECT_ID}-${KUBERNETES_CLUSTER_PREFIX}-training-docs
Verify the document embedder job completed successfully:
kubectl get job -n ${DB_NAMESPACE}
The output is similar to the following:
NAME COMPLETIONS DURATION AGE docs-embedder1716570453361446 1/1 32s 71s
Get the external IP address of the load balancer:
export EXTERNAL_IP=$(kubectl -n ${DB_NAMESPACE} get svc chatbot --output jsonpath='{.status.loadBalancer.ingress[0].ip}') echo http://${EXTERNAL_IP}:80
Open the external IP address in your web browser:
http://EXTERNAL_IP
The chatbot responds with a message similar to the following:
How can I help you?
Ask questions about the content of the uploaded documents. If the chatbot cannot find anything, it answers
I don't know
. For example, you could ask the following:You: Hi, what are Google plans for the future?
An example output from the chatbot is similar to the following:
Bot: Google intends to run on carbon-free energy everywhere, at all times by 2030. To achieve this, it will rely on a combination of renewable energy sources, such as wind and solar, and carbon-free technologies, such as battery storage.
Ask the chatbot a question that is out of context of the uploaded document. For example, you could ask the following:
You: What are Google plans to colonize Mars?
An example output from the chatbot is similar to the following:
Bot: I don't know. The provided context does not mention anything about Google's plans to colonize Mars.
About the application code
This section explains how the application code works. There are three scripts inside the Docker images:
endpoint.py
: receives Eventarc events on each document upload and starts the Kubernetes Jobs to process them.embedding-job.py
: downloads documents from the bucket, creates embeddings, and insert embeddings into the vector database.chat.py
: runs queries over the content of stored documents.
The diagram shows the process of generating answers using the documents data:
In the diagram, the application loads a PDF file, splits the file into chunks, then vectors, then sends the vectors to a vector database. Later, a user asks a question to the chatbot. The RAG chain uses semantic search to search the vector database, then returns the context along with the question to the LLM. The LLM answers the question, and stores the question into chat history.
About endpoint.py
This file processes messages from Eventarc, creates a Kubernetes Job for embedding the document, and accepts request from anywhere on port 5001
Qdrant
Elasticsearch
PGVector
Weaviate
About embedding-job.py
This file processes documents and sends them to the vector database.
Qdrant
Elasticsearch
PGVector
Weaviate
About chat.py
This file configures the model to answer questions using only the provided
context and previous answers. If the context or conversation history does not
match any data, the model returns I don't know
.
Qdrant
Elasticsearch
PGVector
Weaviate
Clean up
To avoid incurring charges to your Google Cloud account for the resources used in this tutorial, either delete the project that contains the resources, or keep the project and delete the individual resources.
Delete the project
The easiest way to avoid billing is to delete the project you created for this tutorial.
Delete a Google Cloud project:
gcloud projects delete PROJECT_ID
If you deleted the project, your clean up is complete. If you didn't delete the project, proceed to delete the individual resources.
Delete individual resources
Delete the Artifact Registry repository:
gcloud artifacts repositories delete ${KUBERNETES_CLUSTER_PREFIX}-images \ --location=${REGION} \ --async
When prompted, type
y
.Delete the Cloud Storage bucket and the Eventarc trigger:
export GOOGLE_OAUTH_ACCESS_TOKEN=$(gcloud auth print-access-token) terraform -chdir=vector-database/terraform/cloud-storage destroy \ -var project_id=${PROJECT_ID} \ -var region=${REGION} \ -var cluster_prefix=${KUBERNETES_CLUSTER_PREFIX} \ -var db_namespace=${DB_NAMESPACE}
When prompted, type
yes
.Eventarc requires that you have a valid endpoint target both during creation and during deletion.
What's next
- Learn about the best practices for deploying databases on GKE.
- Discover solutions for running data-intensive workloads with GKE.