Créer un chatbot RAG avec GKE et Cloud Storage

Autopilot Standard

Ce tutoriel explique comment intégrer une application de grand modèle de langage (LLM) basée sur la génération augmentée par récupération (RAG) avec des fichiers PDF que vous importez dans un bucket Cloud Storage.

Ce guide utilise une base de données comme moteur de recherche sémantique et de stockage qui contient les représentations (embeddings) des documents importés. Vous utiliserez le framework Langchain pour interagir avec les embeddings et les modèles Gemini disponibles via Vertex AI.

Langchain est un framework Python Open Source populaire qui simplifie de nombreuses tâches de machine learning et dispose d'interfaces permettant de s'intégrer à différentes bases de données vectorielles et services d'IA.

Ce tutoriel est destiné aux administrateurs et architectes de plate-forme cloud, aux ingénieurs en ML et aux professionnels du MLOps (DevOps) qui souhaitent déployer des applications RAG LLM sur GKE et Cloud Storage.

Objectifs

Dans ce tutoriel, vous allez apprendre à effectuer les opérations suivantes :

Créez et déployez une application pour créer et stocker des embeddings de documents dans une base de données vectorielle.
Automatisez l'application pour déclencher l'importation de nouveaux documents dans un bucket Cloud Storage.
Déployez une application de chatbot qui utilise la recherche sémantique pour répondre aux questions en fonction du contenu du document.

Architecture de déploiement

Dans ce tutoriel, vous allez créer un bucket Cloud Storage, un déclencheur Eventarc et les services suivants :

embed-docs : Eventarc déclenche ce service chaque fois qu'un utilisateur importe un nouveau document dans le bucket Cloud Storage. Le service démarre un Job Kubernetes qui crée des représentations vectorielles continues pour le document importé et les insère dans une base de données vectorielle.
chatbot : ce service répond aux questions en langage naturel concernant les documents importés à l'aide de la recherche sémantique et de l'API Gemini.

Le schéma suivant illustre le processus d'importation et de vectorisation des documents :

Dans le schéma, l'utilisateur importe des fichiers dans le bucket Cloud Storage. Eventarc s'abonne aux événements metadataUpdated d'objets pour le bucket et utilise le redirecteur d'événements d'Eventarc, qui est une charge de travail Kubernetes, pour appeler le service embed-docs lorsque vous importez un nouveau document. Le service crée ensuite des embeddings pour le document importé. Le service embed-docs stocke les embeddings dans une base de données vectorielle à l'aide du modèle d'embedding Vertex AI.

Le schéma suivant illustre le processus de recherche d'informations sur le contenu du document importé à l'aide du service chatbot :

Les utilisateurs peuvent poser des questions en langage naturel, et le chatbot génère des réponses basées uniquement sur le contenu des fichiers importés. Le chatbot récupère le contexte de la base de données vectorielle à l'aide de la recherche sémantique, puis envoie la question et le contexte à Gemini.

Coûts

Dans ce document, vous utilisez les composants facturables suivants de Google Cloud :

Obtenez une estimation des coûts en fonction de votre utilisation prévue à l'aide du simulateur de coût. Les nouveaux utilisateurs de Google Cloud peuvent bénéficier d'un essai gratuit.

Une fois que vous avez terminé les tâches décrites dans ce document, vous pouvez éviter de continuer à payer des frais en supprimant les ressources que vous avez créées. Pour en savoir plus, consultez la section Effectuer un nettoyage.

Avant de commencer

Dans ce tutoriel, vous utilisez Cloud Shell pour exécuter des commandes. Cloud Shell est un environnement shell permettant de gérer les ressources hébergées sur Google Cloud. Cloud Shell est préinstallé avec les outils de ligne de commande Google Cloud CLI, kubectl et Terraform. Si vous n'utilisez pas Cloud Shell, installez 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

Note: If you installed the gcloud CLI previously, make sure you have the latest version by running

gcloud components
      update

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

Note: If you installed the gcloud CLI previously, make sure you have the latest version by running

gcloud components
      update

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: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.

Créer un cluster

Créez un cluster Qdrant, Elasticsearch ou Postgres :

Qdrant

Suivez les instructions de la section Déployer une base de données vectorielle Qdrant sur GKE pour créer un cluster Qdrant exécuté sur un cluster GKE en mode Autopilot ou Standard.

Elasticsearch

Suivez les instructions de la section Déployer une base de données vectorielle Elasticsearch sur GKE pour créer un cluster Elasticsearch exécuté sur un cluster GKE en mode Autopilot ou Standard.

PGVector

Suivez les instructions de la section Déployer une base de données vectorielle PostgreSQL sur GKE pour créer un cluster PostgreSQL avec PGVector s'exécutant sur un cluster GKE en mode Autopilot ou Standard.

Weaviate

Suivez les instructions de la section Déployer une base de données vectorielle Weaviate sur GKE pour créer un cluster Weaviate s'exécutant sur un cluster GKE en mode Autopilot ou Standard.

Configurer votre environnement

Configurez votre environnement avec Cloud Shell :

Définissez les variables d'environnement pour votre projet :

Qdrant

export PROJECT_ID=PROJECT_ID
export KUBERNETES_CLUSTER_PREFIX=qdrant
export REGION=us-central1
export DB_NAMESPACE=qdrant

Remplacez PROJECT_ID par l'ID de votre projet Google Cloud.

Elasticsearch

export PROJECT_ID=PROJECT_ID
export KUBERNETES_CLUSTER_PREFIX=elasticsearch
export REGION=us-central1
export DB_NAMESPACE=elastic

Remplacez PROJECT_ID par l'ID de votre projet Google Cloud.

PGVector

export PROJECT_ID=PROJECT_ID
export KUBERNETES_CLUSTER_PREFIX=postgres
export REGION=us-central1
export DB_NAMESPACE=pg-ns

Remplacez PROJECT_ID par l'ID de votre projet Google Cloud.

Weaviate

export PROJECT_ID=PROJECT_ID
export KUBERNETES_CLUSTER_PREFIX=weaviate
export REGION=us-central1
export DB_NAMESPACE=weaviate

Remplacez PROJECT_ID par l'ID de votre projet Google Cloud.

Vérifier que votre cluster GKE est en cours d'exécution :

gcloud container clusters list --project=${PROJECT_ID} --region=${REGION}

Le résultat ressemble à ce qui suit :

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

Clonez l'exemple de dépôt de code depuis GitHub :

git clone https://github.com/GoogleCloudPlatform/kubernetes-engine-samples

Accédez au répertoire databases :
```
cd kubernetes-engine-samples/databases
```

Préparer votre infrastructure

Créez un dépôt Artifact Registry, compilez des images Docker et transférez-les vers Artifact Registry :

Créer un dépôt Artifact Registry :

gcloud artifacts repositories create ${KUBERNETES_CLUSTER_PREFIX}-images \
    --repository-format=docker \
    --location=${REGION} \
    --description="Vector database images repository" \
    --async

Définissez les autorisations storage.objectAdmin et artifactregistry.admin sur le compte de service Compute Engine pour utiliser Cloud Build afin de compiler et de transférer des images Docker pour les services embed-docs et chatbot.

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"

Remplacez PROJECT_NUMBER par votre numéro de projet Google Cloud.

Compilez des images Docker pour les services embed-docs et chatbot. L'image embed-docs contient le code Python de l'application qui reçoit les requêtes de redirection Eventarc et le job d'intégration.

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

Vérifiez les images :

gcloud artifacts docker images list $DOCKER_REPO \
    --project=$PROJECT_ID \
    --format="value(IMAGE)"

Le résultat ressemble à ce qui suit :

$REGION-docker.pkg.dev/$PROJECT_ID/${KUBERNETES_CLUSTER_PREFIX}-images/chatbot
$REGION-docker.pkg.dev/$PROJECT_ID/${KUBERNETES_CLUSTER_PREFIX}-images/embed-docs

Déployez un compte de service Kubernetes avec des autorisations pour exécuter des jobs Kubernetes :

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 -

Lorsque vous utilisez Terraform pour créer le cluster GKE et que create_service_account est défini sur "true", un compte de service distinct est créé et utilisé par le cluster et les nœuds. Attribuez le rôle artifactregistry.serviceAgent à ce compte de service Compute Engine pour permettre aux nœuds d'extraire l'image du registre Artifact Registry créé pour embed-docs et chatbot.
```
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"
```
Sans accorder l'accès au compte de service, vos nœuds peuvent rencontrer des problèmes d'autorisation lors de la tentative de récupération de l'image à partir d'Artifact Registry lors du déploiement des services embed-docs et chatbot.

Déployez un déploiement Kubernetes pour les services embed-docs et chatbot :

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 -

Activez les déclencheurs Eventarc pour GKE :
```
gcloud eventarc gke-destinations init
```
Lorsque vous y êtes invité, saisissez y.
Déployez le bucket Cloud Storage et créez un déclencheur Eventarc à l'aide de 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}
```
Lorsque vous y êtes invité, saisissez yes. L'exécution de la commande peut prendre plusieurs minutes.

Terraform crée les ressources suivantes :
- Un bucket Cloud Storage pour importer les documents
- Un déclencheur Eventarc
- Un compte de service Google Cloud nommé service_account_eventarc_name avec l'autorisation d'utiliser Eventarc
- Un compte de service Google Cloud nommé service_account_bucket_name avec l'autorisation de lire le bucket et d'accéder aux modèles Vertex AI.
Le résultat ressemble à ce qui suit :
```
... # Several lines of output omitted

Apply complete! Resources: 15 added, 0 changed, 0 destroyed.

... # Several lines of output omitted
```

Charger des documents et exécuter des requêtes de chatbot

Importez les documents de démonstration et exécutez des requêtes de recherche sur ces documents à l'aide du chatbot :

Importez l'exemple de document carbon-free-energy.pdf dans le bucket :

gsutil cp vector-database/documents/carbon-free-energy.pdf gs://${PROJECT_ID}-${KUBERNETES_CLUSTER_PREFIX}-training-docs

Vérifiez que le job d'intégration de documents a bien été effectué :

kubectl get job -n ${DB_NAMESPACE}

Le résultat ressemble à ce qui suit :

NAME                            COMPLETIONS   DURATION   AGE
docs-embedder1716570453361446   1/1           32s        71s

Obtenez l'adresse IP externe de l'équilibreur de charge :

export EXTERNAL_IP=$(kubectl -n ${DB_NAMESPACE} get svc chatbot --output jsonpath='{.status.loadBalancer.ingress[0].ip}')
echo http://${EXTERNAL_IP}:80

Ouvrez l'adresse IP externe dans votre navigateur :
```
http://EXTERNAL_IP
```
Le chatbot répond par un message semblable à celui-ci :
```
How can I help you?
```

Posez des questions sur le contenu des documents importés. Si le chatbot ne trouve rien, il répond I don't know. Par exemple, vous pouvez poser les questions suivantes :

You: Hi, what are Google plans for the future?

Voici un exemple de sortie du chatbot :

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.

Posez une question au chatbot qui n'est pas en rapport avec le document importé. Par exemple, vous pouvez demander :
```
You: What are Google plans to colonize Mars?
```
Voici un exemple de sortie du chatbot :
```
Bot: I don't know. The provided context does not mention anything about Google's plans to colonize Mars.
```

À propos du code d'application

Cette section explique le fonctionnement du code. Les images Docker contiennent trois scripts :

endpoint.py : reçoit les événements Eventarc lors de chaque importation de document et lance les jobs Kubernetes pour les traiter.
embedding-job.py : télécharge les documents du bucket, crée des embeddings et les insère dans la base de données de vecteurs.
chat.py : exécute des requêtes sur le contenu des documents stockés.

Le schéma illustre le processus de génération des réponses à l'aide des données des documents :

Dans le schéma, l'application charge un fichier PDF, le divise en segments, puis en vecteurs, avant d'envoyer les vecteurs à une base de données vectorielle. Plus tard, un utilisateur pose une question au chatbot. La chaîne RAG utilise la recherche sémantique pour rechercher dans la base de données vectorielle, puis renvoie le contexte avec la question au LLM. Le LLM répond à la question et la stocke dans l'historique du chat.

À propos de `endpoint.py`

Ce fichier traite les messages d'Eventarc, crée un job Kubernetes pour intégrer le document et accepte les requêtes depuis n'importe où sur le port 5001.

Qdrant

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from flask import Flask, jsonify
from flask import request
import logging
import sys,os, time
from kubernetes import client, config, utils
import kubernetes.client
from kubernetes.client.rest import ApiException


app = Flask(__name__)
@app.route('/check')
def message():
    return jsonify({"Message": "Hi there"})


@app.route('/', methods=['POST'])
def bucket():
    request_data = request.get_json()
    print(request_data)
    bckt = request_data['bucket']
    f_name = request_data['name']
    id = request_data['generation'] 
    kube_create_job(bckt, f_name, id)
    return "ok"

# Set logging
logging.basicConfig(stream=sys.stdout, level=logging.INFO)

# Setup K8 configs
config.load_incluster_config()
def kube_create_job_object(name, container_image, bucket_name, f_name, namespace="qdrant", container_name="jobcontainer", env_vars={}):

    body = client.V1Job(api_version="batch/v1", kind="Job")
    body.metadata = client.V1ObjectMeta(namespace=namespace, name=name)
    body.status = client.V1JobStatus()

    template = client.V1PodTemplate()
    template.template = client.V1PodTemplateSpec()
    env_list = [
        client.V1EnvVar(name="QDRANT_URL", value=os.getenv("QDRANT_URL")),
        client.V1EnvVar(name="COLLECTION_NAME", value="training-docs"), 
        client.V1EnvVar(name="FILE_NAME", value=f_name), 
        client.V1EnvVar(name="BUCKET_NAME", value=bucket_name),
        client.V1EnvVar(name="APIKEY", value_from=client.V1EnvVarSource(secret_key_ref=client.V1SecretKeySelector(key="api-key", name="qdrant-database-apikey"))), 
    ]

    container = client.V1Container(name=container_name, image=container_image, env=env_list)
    template.template.spec = client.V1PodSpec(containers=[container], restart_policy='Never', service_account='embed-docs-sa')

    body.spec = client.V1JobSpec(backoff_limit=3, ttl_seconds_after_finished=60, template=template.template)
    return body
def kube_test_credentials():
    try: 
        api_response = api_instance.get_api_resources()
        logging.info(api_response)
    except ApiException as e:
        print("Exception when calling API: %s\n" % e)

def kube_create_job(bckt, f_name, id):
    container_image = os.getenv("JOB_IMAGE")
    namespace = os.getenv("JOB_NAMESPACE")
    name = "docs-embedder" + id
    body = kube_create_job_object(name, container_image, bckt, f_name)
    v1=client.BatchV1Api()
    try: 
        v1.create_namespaced_job(namespace, body, pretty=True)
    except ApiException as e:
        print("Exception when calling BatchV1Api->create_namespaced_job: %s\n" % e)
    return

if __name__ == '__main__':
    app.run('0.0.0.0', port=5001, debug=True)

Elasticsearch

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from flask import Flask, jsonify
from flask import request
import logging
import sys,os, time
from kubernetes import client, config, utils
import kubernetes.client
from kubernetes.client.rest import ApiException


app = Flask(__name__)
@app.route('/check')
def message():
    return jsonify({"Message": "Hi there"})


@app.route('/', methods=['POST'])
def bucket():
    request_data = request.get_json()
    print(request_data)
    bckt = request_data['bucket']
    f_name = request_data['name']
    id = request_data['generation'] 
    kube_create_job(bckt, f_name, id)
    return "ok"

# Set logging
logging.basicConfig(stream=sys.stdout, level=logging.INFO)

# Setup K8 configs
config.load_incluster_config()

def kube_create_job_object(name, container_image, bucket_name, f_name, namespace="elastic", container_name="jobcontainer", env_vars={}):

    body = client.V1Job(api_version="batch/v1", kind="Job")
    body.metadata = client.V1ObjectMeta(namespace=namespace, name=name)
    body.status = client.V1JobStatus()

    template = client.V1PodTemplate()
    template.template = client.V1PodTemplateSpec()
    env_list = [
        client.V1EnvVar(name="ES_URL", value=os.getenv("ES_URL")),
        client.V1EnvVar(name="INDEX_NAME", value="training-docs"), 
        client.V1EnvVar(name="FILE_NAME", value=f_name), 
        client.V1EnvVar(name="BUCKET_NAME", value=bucket_name),
        client.V1EnvVar(name="PASSWORD", value_from=client.V1EnvVarSource(secret_key_ref=client.V1SecretKeySelector(key="elastic", name="elasticsearch-ha-es-elastic-user"))), 
    ]

    container = client.V1Container(name=container_name, image=container_image, image_pull_policy='Always', env=env_list)
    template.template.spec = client.V1PodSpec(containers=[container], restart_policy='Never', service_account='embed-docs-sa')

    body.spec = client.V1JobSpec(backoff_limit=3, ttl_seconds_after_finished=60, template=template.template)
    return body

def kube_test_credentials():
    try: 
        api_response = api_instance.get_api_resources()
        logging.info(api_response)
    except ApiException as e:
        print("Exception when calling API: %s\n" % e)

def kube_create_job(bckt, f_name, id):
    container_image = os.getenv("JOB_IMAGE")
    namespace = os.getenv("JOB_NAMESPACE")
    name = "docs-embedder" + id
    body = kube_create_job_object(name, container_image, bckt, f_name)
    v1=client.BatchV1Api()
    try: 
        v1.create_namespaced_job(namespace, body, pretty=True)
    except ApiException as e:
        print("Exception when calling BatchV1Api->create_namespaced_job: %s\n" % e)
    return

if __name__ == '__main__':
    app.run('0.0.0.0', port=5001, debug=True)

PGVector

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from flask import Flask, jsonify
from flask import request
import logging
import sys,os, time
from kubernetes import client, config, utils
import kubernetes.client
from kubernetes.client.rest import ApiException


app = Flask(__name__)
@app.route('/check')
def message():
    return jsonify({"Message": "Hi there"})


@app.route('/', methods=['POST'])
def bucket():
    request_data = request.get_json()
    print(request_data)
    bckt = request_data['bucket']
    f_name = request_data['name']
    id = request_data['generation'] 
    kube_create_job(bckt, f_name, id)
    return "ok"

# Set logging
logging.basicConfig(stream=sys.stdout, level=logging.INFO)

# Setup K8 configs
config.load_incluster_config()
def kube_create_job_object(name, container_image, bucket_name, f_name, namespace="pg-ns", container_name="jobcontainer", env_vars={}):

    body = client.V1Job(api_version="batch/v1", kind="Job")
    body.metadata = client.V1ObjectMeta(namespace=namespace, name=name)
    body.status = client.V1JobStatus()

    template = client.V1PodTemplate()
    template.template = client.V1PodTemplateSpec()
    env_list = [
        client.V1EnvVar(name="POSTGRES_HOST", value=os.getenv("POSTGRES_HOST")),
        client.V1EnvVar(name="DATABASE_NAME", value="app"), 
        client.V1EnvVar(name="COLLECTION_NAME", value="training-docs"), 
        client.V1EnvVar(name="FILE_NAME", value=f_name), 
        client.V1EnvVar(name="BUCKET_NAME", value=bucket_name),
        client.V1EnvVar(name="PASSWORD", value_from=client.V1EnvVarSource(secret_key_ref=client.V1SecretKeySelector(key="password", name="gke-pg-cluster-app"))), 
        client.V1EnvVar(name="USERNAME", value_from=client.V1EnvVarSource(secret_key_ref=client.V1SecretKeySelector(key="username", name="gke-pg-cluster-app"))), 
    ]

    container = client.V1Container(name=container_name, image=container_image, image_pull_policy='Always', env=env_list)
    template.template.spec = client.V1PodSpec(containers=[container], restart_policy='Never', service_account='embed-docs-sa')

    body.spec = client.V1JobSpec(backoff_limit=3, ttl_seconds_after_finished=60, template=template.template)
    return body
def kube_test_credentials():
    try: 
        api_response = api_instance.get_api_resources()
        logging.info(api_response)
    except ApiException as e:
        print("Exception when calling API: %s\n" % e)

def kube_create_job(bckt, f_name, id):
    container_image = os.getenv("JOB_IMAGE")
    namespace = os.getenv("JOB_NAMESPACE")
    name = "docs-embedder" + id
    body = kube_create_job_object(name, container_image, bckt, f_name)
    v1=client.BatchV1Api()
    try: 
        v1.create_namespaced_job(namespace, body, pretty=True)
    except ApiException as e:
        print("Exception when calling BatchV1Api->create_namespaced_job: %s\n" % e)
    return

if __name__ == '__main__':
    app.run('0.0.0.0', port=5001, debug=True)

Weaviate

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from flask import Flask, jsonify
from flask import request
import logging
import sys,os, time
from kubernetes import client, config, utils
import kubernetes.client
from kubernetes.client.rest import ApiException


app = Flask(__name__)
@app.route('/check')
def message():
    return jsonify({"Message": "Hi there"})


@app.route('/', methods=['POST'])
def bucket():
    request_data = request.get_json()
    print(request_data)
    bckt = request_data['bucket']
    f_name = request_data['name']
    id = request_data['generation'] 
    kube_create_job(bckt, f_name, id)
    return "ok"

# Set logging
logging.basicConfig(stream=sys.stdout, level=logging.INFO)

# Setup K8 configs
config.load_incluster_config()
def kube_create_job_object(name, container_image, bucket_name, f_name, namespace, container_name="jobcontainer", env_vars={}):

    body = client.V1Job(api_version="batch/v1", kind="Job")
    body.metadata = client.V1ObjectMeta(namespace=namespace, name=name)
    body.status = client.V1JobStatus()

    template = client.V1PodTemplate()
    template.template = client.V1PodTemplateSpec()
    env_list = [
        client.V1EnvVar(name="WEAVIATE_ENDPOINT", value=os.getenv("WEAVIATE_ENDPOINT")),
        client.V1EnvVar(name="WEAVIATE_GRPC_ENDPOINT", value=os.getenv("WEAVIATE_GRPC_ENDPOINT")),
        client.V1EnvVar(name="FILE_NAME", value=f_name), 
        client.V1EnvVar(name="BUCKET_NAME", value=bucket_name),
        client.V1EnvVar(name="APIKEY", value_from=client.V1EnvVarSource(secret_key_ref=client.V1SecretKeySelector(key="AUTHENTICATION_APIKEY_ALLOWED_KEYS", name="apikeys"))), 
    ]

    container = client.V1Container(name=container_name, image=container_image, image_pull_policy='Always', env=env_list)
    template.template.spec = client.V1PodSpec(containers=[container], restart_policy='Never', service_account='embed-docs-sa')

    body.spec = client.V1JobSpec(backoff_limit=3, ttl_seconds_after_finished=60, template=template.template)
    return body
def kube_test_credentials():
    try: 
        api_response = api_instance.get_api_resources()
        logging.info(api_response)
    except ApiException as e:
        print("Exception when calling API: %s\n" % e)

def kube_create_job(bckt, f_name, id):
    container_image = os.getenv("JOB_IMAGE")
    namespace = os.getenv("JOB_NAMESPACE")
    name = "docs-embedder" + id
    body = kube_create_job_object(name, container_image, bckt, f_name, namespace)
    v1=client.BatchV1Api()
    try: 
        v1.create_namespaced_job(namespace, body, pretty=True)
    except ApiException as e:
        print("Exception when calling BatchV1Api->create_namespaced_job: %s\n" % e)
    return

if __name__ == '__main__':
    app.run('0.0.0.0', port=5001, debug=True)

À propos de `embedding-job.py`

Ce fichier traite les documents et les envoie à la base de données de vecteurs.

Qdrant

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from langchain_google_vertexai import ChatVertexAI
from langchain.prompts import ChatPromptTemplate
from langchain_google_vertexai import VertexAIEmbeddings
from langchain.memory import ConversationBufferWindowMemory
from langchain_community.vectorstores import Qdrant
from qdrant_client import QdrantClient
import streamlit as st
import os

vertexAI = ChatVertexAI(model_name="gemini-pro", streaming=True, convert_system_message_to_human=True)
prompt_template = ChatPromptTemplate.from_messages(
    [
        ("system", "You are a helpful assistant who helps in finding answers to questions using the provided context."),
        ("human", """
        The answer should be based on the text context given in "text_context" and the conversation history given in "conversation_history" along with its Caption: \n
        Base your response on the provided text context and the current conversation history to answer the query.
        Select the most relevant information from the context.
        Generate a draft response using the selected information. Remove duplicate content from the draft response.
        Generate your final response after adjusting it to increase accuracy and relevance.
        Now only show your final response!
        If you do not know the answer or context is not relevant, response with "I don't know".

        text_context:
        {context}

        conversation_history:
        {history}

        query:
        {query}
        """),
    ]
)

embedding_model = VertexAIEmbeddings("textembedding-gecko@001")

client = QdrantClient(
    url=os.getenv("QDRANT_URL"),
    api_key=os.getenv("APIKEY"),
)
collection_name = os.getenv("COLLECTION_NAME")
vector_search = Qdrant(client, collection_name, embeddings=embedding_model)
def format_docs(docs):
    return "\n\n".join([d.page_content for d in docs])

st.title("🤖 Chatbot")
if "messages" not in st.session_state:
    st.session_state["messages"] = [{"role": "ai", "content": "How can I help you?"}]
if "memory" not in st.session_state:
    st.session_state["memory"] = ConversationBufferWindowMemory(
        memory_key="history",
        ai_prefix="Bob",
        human_prefix="User",
        k=3,
    )
for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.write(message["content"])
if chat_input := st.chat_input():
    with st.chat_message("human"):
        st.write(chat_input)
        st.session_state.messages.append({"role": "human", "content": chat_input})

    found_docs = vector_search.similarity_search(chat_input)
    context = format_docs(found_docs)

    prompt_value = prompt_template.format_messages(name="Bob", query=chat_input, context=context, history=st.session_state.memory.load_memory_variables({}))
    with st.chat_message("ai"):
        with st.spinner("Typing..."):
            content = ""
            with st.empty():
                for chunk in vertexAI.stream(prompt_value):
                    content += chunk.content
                    st.write(content)
            st.session_state.messages.append({"role": "ai", "content": content})

    st.session_state.memory.save_context({"input": chat_input}, {"output": content})

Elasticsearch

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from langchain_google_vertexai import VertexAIEmbeddings
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from elasticsearch import Elasticsearch
from langchain_community.vectorstores.elasticsearch import ElasticsearchStore
from google.cloud import storage
import os

bucketname = os.getenv("BUCKET_NAME")
filename = os.getenv("FILE_NAME")

storage_client = storage.Client()
bucket = storage_client.bucket(bucketname)
blob = bucket.blob(filename)
blob.download_to_filename("/documents/" + filename)

loader = PyPDFLoader("/documents/" + filename)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
documents = loader.load_and_split(text_splitter)

embeddings = VertexAIEmbeddings("textembedding-gecko@001")

client = Elasticsearch(
    [os.getenv("ES_URL")], 
    verify_certs=False, 
    ssl_show_warn=False,
    basic_auth=("elastic", os.getenv("PASSWORD"))
)

db = ElasticsearchStore.from_documents(
    documents,
    embeddings,
    es_connection=client,
    index_name=os.getenv("INDEX_NAME")
)
db.client.indices.refresh(index=os.getenv("INDEX_NAME"))

print(filename + " was successfully embedded") 
print(f"# of vectors = {len(documents)}")

PGVector

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from langchain_google_vertexai import VertexAIEmbeddings
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores.pgvector import PGVector
from google.cloud import storage
import os
bucketname = os.getenv("BUCKET_NAME")
filename = os.getenv("FILE_NAME")

storage_client = storage.Client()
bucket = storage_client.bucket(bucketname)
blob = bucket.blob(filename)
blob.download_to_filename("/documents/" + filename)

loader = PyPDFLoader("/documents/" + filename)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
documents = loader.load_and_split(text_splitter)
for document in documents:
    document.page_content = document.page_content.replace('\x00', '')

embeddings = VertexAIEmbeddings("textembedding-gecko@001")

CONNECTION_STRING = PGVector.connection_string_from_db_params(
    driver="psycopg2",
    host=os.environ.get("POSTGRES_HOST"),
    port=5432,
    database=os.environ.get("DATABASE_NAME"),
    user=os.environ.get("USERNAME"),
    password=os.environ.get("PASSWORD"),
)
COLLECTION_NAME = os.environ.get("COLLECTION_NAME")

db = PGVector.from_documents(
    embedding=embeddings,
    documents=documents,
    collection_name=COLLECTION_NAME,
    connection_string=CONNECTION_STRING,
    use_jsonb=True
)

print(filename + " was successfully embedded") 
print(f"# of vectors = {len(documents)}")

Weaviate

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from langchain_google_vertexai import VertexAIEmbeddings
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
import weaviate
from weaviate.connect import ConnectionParams
from langchain_weaviate.vectorstores import WeaviateVectorStore
from google.cloud import storage
import os
bucketname = os.getenv("BUCKET_NAME")
filename = os.getenv("FILE_NAME")

storage_client = storage.Client()
bucket = storage_client.bucket(bucketname)
blob = bucket.blob(filename)
blob.download_to_filename("/documents/" + filename)

loader = PyPDFLoader("/documents/" + filename)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
documents = loader.load_and_split(text_splitter)

embeddings = VertexAIEmbeddings("textembedding-gecko@001")

auth_config = weaviate.auth.AuthApiKey(api_key=os.getenv("APIKEY"))
client = weaviate.WeaviateClient(
    connection_params=ConnectionParams.from_params(
        http_host=os.getenv("WEAVIATE_ENDPOINT"),
        http_port="80",
        http_secure=False,
        grpc_host=os.getenv("WEAVIATE_GRPC_ENDPOINT"),
        grpc_port="50051",
        grpc_secure=False,
    ),
    auth_client_secret=auth_config
)
client.connect()
if not client.collections.exists("trainingdocs"):
    collection = client.collections.create(name="trainingdocs")
db = WeaviateVectorStore.from_documents(documents, embeddings, client=client, index_name="trainingdocs")

print(filename + " was successfully embedded") 
print(f"# of vectors = {len(documents)}")

À propos de `chat.py`

Ce fichier configure le modèle pour qu'il réponde aux questions en utilisant uniquement le contexte fourni et les réponses précédentes. Si le contexte ou l'historique des conversations ne correspond à aucune donnée, le modèle renvoie I don't know.

Qdrant

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from flask import Flask, jsonify
from flask import request
import logging
import sys,os, time
from kubernetes import client, config, utils
import kubernetes.client
from kubernetes.client.rest import ApiException


app = Flask(__name__)
@app.route('/check')
def message():
    return jsonify({"Message": "Hi there"})


@app.route('/', methods=['POST'])
def bucket():
    request_data = request.get_json()
    print(request_data)
    bckt = request_data['bucket']
    f_name = request_data['name']
    id = request_data['generation'] 
    kube_create_job(bckt, f_name, id)
    return "ok"

# Set logging
logging.basicConfig(stream=sys.stdout, level=logging.INFO)

# Setup K8 configs
config.load_incluster_config()
def kube_create_job_object(name, container_image, bucket_name, f_name, namespace="qdrant", container_name="jobcontainer", env_vars={}):

    body = client.V1Job(api_version="batch/v1", kind="Job")
    body.metadata = client.V1ObjectMeta(namespace=namespace, name=name)
    body.status = client.V1JobStatus()

    template = client.V1PodTemplate()
    template.template = client.V1PodTemplateSpec()
    env_list = [
        client.V1EnvVar(name="QDRANT_URL", value=os.getenv("QDRANT_URL")),
        client.V1EnvVar(name="COLLECTION_NAME", value="training-docs"), 
        client.V1EnvVar(name="FILE_NAME", value=f_name), 
        client.V1EnvVar(name="BUCKET_NAME", value=bucket_name),
        client.V1EnvVar(name="APIKEY", value_from=client.V1EnvVarSource(secret_key_ref=client.V1SecretKeySelector(key="api-key", name="qdrant-database-apikey"))), 
    ]

    container = client.V1Container(name=container_name, image=container_image, env=env_list)
    template.template.spec = client.V1PodSpec(containers=[container], restart_policy='Never', service_account='embed-docs-sa')

    body.spec = client.V1JobSpec(backoff_limit=3, ttl_seconds_after_finished=60, template=template.template)
    return body
def kube_test_credentials():
    try: 
        api_response = api_instance.get_api_resources()
        logging.info(api_response)
    except ApiException as e:
        print("Exception when calling API: %s\n" % e)

def kube_create_job(bckt, f_name, id):
    container_image = os.getenv("JOB_IMAGE")
    namespace = os.getenv("JOB_NAMESPACE")
    name = "docs-embedder" + id
    body = kube_create_job_object(name, container_image, bckt, f_name)
    v1=client.BatchV1Api()
    try: 
        v1.create_namespaced_job(namespace, body, pretty=True)
    except ApiException as e:
        print("Exception when calling BatchV1Api->create_namespaced_job: %s\n" % e)
    return

if __name__ == '__main__':
    app.run('0.0.0.0', port=5001, debug=True)

Elasticsearch

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from langchain_google_vertexai import ChatVertexAI
from langchain.prompts import ChatPromptTemplate
from langchain_google_vertexai import VertexAIEmbeddings
from langchain.memory import ConversationBufferWindowMemory
from elasticsearch import Elasticsearch
from langchain_community.vectorstores.elasticsearch import ElasticsearchStore
import streamlit as st
import os

vertexAI = ChatVertexAI(model_name="gemini-pro", streaming=True, convert_system_message_to_human=True)
prompt_template = ChatPromptTemplate.from_messages(
    [
        ("system", "You are a helpful assistant who helps in finding answers to questions using the provided context."),
        ("human", """
        The answer should be based on the text context given in "text_context" and the conversation history given in "conversation_history" along with its Caption: \n
        Base your response on the provided text context and the current conversation history to answer the query.
        Select the most relevant information from the context.
        Generate a draft response using the selected information. Remove duplicate content from the draft response.
        Generate your final response after adjusting it to increase accuracy and relevance.
        Now only show your final response!
        If you do not know the answer or context is not relevant, response with "I don't know".

        text_context:
        {context}

        conversation_history:
        {history}

        query:
        {query}
        """),
    ]
)

embedding_model = VertexAIEmbeddings("textembedding-gecko@001")

client = Elasticsearch(
    [os.getenv("ES_URL")], 
    verify_certs=False, 
    ssl_show_warn=False,
    basic_auth=("elastic", os.getenv("PASSWORD"))
)
vector_search = ElasticsearchStore(
    index_name=os.getenv("INDEX_NAME"),
    es_connection=client,
    embedding=embedding_model
)

def format_docs(docs):
    return "\n\n".join([d.page_content for d in docs])

st.title("🤖 Chatbot")
if "messages" not in st.session_state:
    st.session_state["messages"] = [{"role": "ai", "content": "How can I help you?"}]

if "memory" not in st.session_state:
    st.session_state["memory"] = ConversationBufferWindowMemory(
        memory_key="history",
        ai_prefix="Bot",
        human_prefix="User",
        k=3,
    )

for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.write(message["content"])

if chat_input := st.chat_input():
    with st.chat_message("human"):
        st.write(chat_input)
        st.session_state.messages.append({"role": "human", "content": chat_input})

    found_docs = vector_search.similarity_search(chat_input)
    context = format_docs(found_docs)

    prompt_value = prompt_template.format_messages(name="Bot", query=chat_input, context=context, history=st.session_state.memory.load_memory_variables({}))
    with st.chat_message("ai"):
        with st.spinner("Typing..."):
            content = ""
            with st.empty():
                for chunk in vertexAI.stream(prompt_value):
                    content += chunk.content
                    st.write(content)
            st.session_state.messages.append({"role": "ai", "content": content})

    st.session_state.memory.save_context({"input": chat_input}, {"output": content})

PGVector

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from langchain_google_vertexai import ChatVertexAI
from langchain.prompts import ChatPromptTemplate
from langchain_google_vertexai import VertexAIEmbeddings
from langchain.memory import ConversationBufferWindowMemory
from langchain_community.vectorstores.pgvector import PGVector
import streamlit as st
import os

vertexAI = ChatVertexAI(model_name="gemini-pro", streaming=True, convert_system_message_to_human=True)
prompt_template = ChatPromptTemplate.from_messages(
    [
        ("system", "You are a helpful assistant who helps in finding answers to questions using the provided context."),
        ("human", """
        The answer should be based on the text context given in "text_context" and the conversation history given in "conversation_history" along with its Caption: \n
        Base your response on the provided text context and the current conversation history to answer the query.
        Select the most relevant information from the context.
        Generate a draft response using the selected information. Remove duplicate content from the draft response.
        Generate your final response after adjusting it to increase accuracy and relevance.
        Now only show your final response!
        If you do not know the answer or context is not relevant, response with "I don't know".

        text_context:
        {context}

        conversation_history:
        {history}

        query:
        {query}
        """),
    ]
)

embedding_model = VertexAIEmbeddings("textembedding-gecko@001")

CONNECTION_STRING = PGVector.connection_string_from_db_params(
    driver="psycopg2",
    host=os.environ.get("POSTGRES_HOST"),
    port=5432,
    database=os.environ.get("DATABASE_NAME"),
    user=os.environ.get("USERNAME"),
    password=os.environ.get("PASSWORD"),
)
COLLECTION_NAME = os.environ.get("COLLECTION_NAME"),

vector_search = PGVector(
    collection_name=COLLECTION_NAME,
    connection_string=CONNECTION_STRING,
    embedding_function=embedding_model,
)

def format_docs(docs):
    return "\n\n".join([d.page_content for d in docs])

st.title("🤖 Chatbot")
if "messages" not in st.session_state:
    st.session_state["messages"] = [{"role": "ai", "content": "How can I help you?"}]

if "memory" not in st.session_state:
    st.session_state["memory"] = ConversationBufferWindowMemory(
        memory_key="history",
        ai_prefix="Bot",
        human_prefix="User",
        k=3,
    )

for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.write(message["content"])

if chat_input := st.chat_input():
    with st.chat_message("human"):
        st.write(chat_input)
        st.session_state.messages.append({"role": "human", "content": chat_input})

    found_docs = vector_search.similarity_search(chat_input)
    context = format_docs(found_docs)

    prompt_value = prompt_template.format_messages(name="Bot", query=chat_input, context=context, history=st.session_state.memory.load_memory_variables({}))
    with st.chat_message("ai"):
        with st.spinner("Typing..."):
            content = ""
            with st.empty():
                for chunk in vertexAI.stream(prompt_value):
                    content += chunk.content
                    st.write(content)
            st.session_state.messages.append({"role": "ai", "content": content})

    st.session_state.memory.save_context({"input": chat_input}, {"output": content})

Weaviate

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from langchain_google_vertexai import ChatVertexAI
from langchain.prompts import ChatPromptTemplate
from langchain_google_vertexai import VertexAIEmbeddings
from langchain.memory import ConversationBufferWindowMemory
import weaviate
from weaviate.connect import ConnectionParams
from langchain_weaviate.vectorstores import WeaviateVectorStore
import streamlit as st
import os

vertexAI = ChatVertexAI(model_name="gemini-pro", streaming=True, convert_system_message_to_human=True)
prompt_template = ChatPromptTemplate.from_messages(
    [
        ("system", "You are a helpful assistant who helps in finding answers to questions using the provided context."),
        ("human", """
        The answer should be based on the text context given in "text_context" and the conversation history given in "conversation_history" along with its Caption: \n
        Base your response on the provided text context and the current conversation history to answer the query.
        Select the most relevant information from the context.
        Generate a draft response using the selected information. Remove duplicate content from the draft response.
        Generate your final response after adjusting it to increase accuracy and relevance.
        Now only show your final response!
        If you do not know the answer or context is not relevant, response with "I don't know".

        text_context:
        {context}

        conversation_history:
        {history}

        query:
        {query}
        """),
    ]
)

embedding_model = VertexAIEmbeddings("textembedding-gecko@001")

auth_config = weaviate.auth.AuthApiKey(api_key=os.getenv("APIKEY"))
client = weaviate.WeaviateClient(
    connection_params=ConnectionParams.from_params(
        http_host=os.getenv("WEAVIATE_ENDPOINT"),
        http_port="80",
        http_secure=False,
        grpc_host=os.getenv("WEAVIATE_GRPC_ENDPOINT"),
        grpc_port="50051",
        grpc_secure=False,
    ),
    auth_client_secret=auth_config
)
client.connect()

vector_search = WeaviateVectorStore.from_documents([],embedding_model,client=client, index_name="trainingdocs")

def format_docs(docs):
    return "\n\n".join([d.page_content for d in docs])

st.title("🤖 Chatbot")
if "messages" not in st.session_state:
    st.session_state["messages"] = [{"role": "ai", "content": "How can I help you?"}]

if "memory" not in st.session_state:
    st.session_state["memory"] = ConversationBufferWindowMemory(
        memory_key="history",
        ai_prefix="Bot",
        human_prefix="User",
        k=3,
    )

for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.write(message["content"])

if chat_input := st.chat_input():
    with st.chat_message("human"):
        st.write(chat_input)
        st.session_state.messages.append({"role": "human", "content": chat_input})

    found_docs = vector_search.similarity_search(chat_input)
    context = format_docs(found_docs)

    prompt_value = prompt_template.format_messages(name="Bot", query=chat_input, context=context, history=st.session_state.memory.load_memory_variables({}))
    with st.chat_message("ai"):
        with st.spinner("Typing..."):
            content = ""
            with st.empty():
                for chunk in vertexAI.stream(prompt_value):
                    content += chunk.content
                    st.write(content)
            st.session_state.messages.append({"role": "ai", "content": content})

    st.session_state.memory.save_context({"input": chat_input}, {"output": content})

Effectuer un nettoyage

Pour éviter que les ressources utilisées lors de ce tutoriel soient facturées sur votre compte Google Cloud, supprimez le projet contenant les ressources, ou conservez le projet et supprimez les ressources individuelles.

Supprimer le projet

Le moyen le plus simple d'empêcher la facturation est de supprimer le projet que vous avez créé pour ce tutoriel.

Attention : La suppression d'un projet entraîne les effets décrits ci-dessous :

Tout le contenu du projet est supprimé. Si vous avez utilisé un projet existant pour les tâches décrites dans ce document et que vous le supprimez, vous supprimerez également tout autre travail effectué dans le projet.
Les ID de projets personnalisés sont perdus. Lorsque vous avez créé ce projet, vous avez peut-être créé un ID de projet personnalisé que vous souhaitez utiliser à l'avenir. Pour conserver les URL qui utilisent l'ID de projet, telle qu'une URL appspot.com, supprimez les ressources sélectionnées dans le projet au lieu de supprimer l'ensemble du projet.

Si vous envisagez d'explorer plusieurs architectures, tutoriels et guides de démarrage rapide, réutiliser des projets peut vous aider à ne pas dépasser les limites de quotas des projets.

Delete a Google Cloud project:

gcloud projects delete PROJECT_ID

Si vous avez supprimé le projet, le nettoyage est terminé. Si vous n'avez pas supprimé le projet, suivez les étapes ci-après afin de supprimer les ressources individuelles.

Supprimer des ressources individuelles

Supprimez le dépôt Artifact Registry :

gcloud artifacts repositories delete ${KUBERNETES_CLUSTER_PREFIX}-images \
    --location=${REGION} \
    --async

Lorsque vous y êtes invité, saisissez y.

Supprimez le bucket Cloud Storage et le déclencheur Eventarc :

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}

Lorsque vous y êtes invité, saisissez yes.

Eventarc nécessite que vous disposiez d'une cible de point de terminaison valide lors de la création et de la suppression.

Étape suivante

Découvrez les bonnes pratiques pour le déploiement de bases de données sur GKE.
Découvrez des solutions pour exécuter des charges de travail utilisant beaucoup de données avec GKE.

Créer un chatbot RAG avec GKE et Cloud Storage

Objectifs

Architecture de déploiement

Coûts

Avant de commencer

Créer un cluster

Qdrant

Elasticsearch

PGVector

Weaviate

Configurer votre environnement

Qdrant

Elasticsearch

PGVector

Weaviate

Préparer votre infrastructure

Qdrant

Elasticsearch

PGVector

Weaviate

Qdrant

Elasticsearch

PGVector

Weaviate

Qdrant

Elasticsearch

PGVector

Weaviate

Charger des documents et exécuter des requêtes de chatbot

À propos du code d'application

À propos de endpoint.py

Qdrant

Elasticsearch

PGVector

Weaviate

À propos de embedding-job.py

Qdrant

Elasticsearch

PGVector

Weaviate

À propos de chat.py

Qdrant

Elasticsearch

PGVector

Weaviate

Effectuer un nettoyage

Supprimer le projet

Supprimer des ressources individuelles

Étape suivante

À propos de `endpoint.py`

À propos de `embedding-job.py`

À propos de `chat.py`