Crea un chatbot de RAG con GKE y Cloud Storage

En este instructivo, se muestra cómo integrar una aplicación de modelo de lenguaje grande (LLM) basada en la generación mejorada por recuperación (RAG) con archivos PDF que subes a un bucket de Cloud Storage.

En esta guía, se usa una base de datos como motor de almacenamiento y búsqueda semántica que contiene las representaciones (incorporaciones) de los documentos subidos. Usas el framework de Langchain para interactuar con las incorporaciones y los modelos de Gemini disponibles a través de Vertex AI.

Langchain es un framework de Python de código abierto popular que simplifica muchas tareas de aprendizaje automático y tiene interfaces para integrarse con diferentes bases de datos vectoriales y servicios de IA.

Este instructivo está dirigido a administradores y arquitectos de plataformas de nube, ingenieros de AA y profesionales de MLOps (DevOps) interesados en implementar aplicaciones de LLM de RAG en GKE y Cloud Storage.

Crea un clúster

Crea un clúster de Qdrant, Elasticsearch o Postgres:

Qdrant

Sigue las instrucciones que se indican en Implementa una base de datos vectorial de Qdrant en GKE para crear un clúster de Qdrant que se ejecute en un clúster de GKE en modo Autopilot o estándar.

Elasticsearch

Sigue las instrucciones que se indican en Implementa una base de datos vectorial de Elasticsearch en GKE para crear un clúster de Elasticsearch que se ejecute en un clúster de GKE en modo Autopilot o Standard.

PGVector

Sigue las instrucciones en Implementa una base de datos vectorial de PostgreSQL en GKE para crear un clúster de Postgres con PGVector que se ejecute en un clúster de GKE en modo Autopilot o Standard.

Weaviate

Sigue las instrucciones para implementar una base de datos vectorial de Weaviate en GKE y crear un clúster de Weaviate que se ejecute en un clúster de GKE en modo Autopilot o Standard.

Configura tu entorno

Configura tu entorno con Cloud Shell:

  1. Configura las variables de entorno de tu proyecto:

    Qdrant

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

    Reemplaza PROJECT_ID por el ID del proyectoGoogle Cloud .

    Elasticsearch

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

    Reemplaza PROJECT_ID por el ID del proyectoGoogle Cloud .

    PGVector

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

    Reemplaza PROJECT_ID por el ID del proyectoGoogle Cloud .

    Weaviate

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

    Reemplaza PROJECT_ID por el ID del proyectoGoogle Cloud .

  2. Verifica que el clúster de GKE esté en ejecución:

    gcloud container clusters list --project=${PROJECT_ID} --location=${CONTROL_PLANE_LOCATION}
    

    El resultado es similar a este:

    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
    
  3. Clona el repositorio de código de muestra de GitHub:

    git clone https://github.com/GoogleCloudPlatform/kubernetes-engine-samples
    
  4. Navega al directorio databases:

    cd kubernetes-engine-samples/databases
    

Prepara tu infraestructura

Crea un repositorio de Artifact Registry, compila imágenes de Docker y envíalas a Artifact Registry:

  1. Crea un repositorio de Artifact Registry:

    gcloud artifacts repositories create ${KUBERNETES_CLUSTER_PREFIX}-images \
        --repository-format=docker \
        --location=${REGION} \
        --description="Vector database images repository" \
        --async
    
  2. Configura los permisos storage.objectAdmin y artifactregistry.admin en la cuenta de servicio de Compute Engine para usar Cloud Build y compilar y enviar imágenes de Docker para los servicios embed-docs y 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"
    

    Reemplaza PROJECT_NUMBER por elGoogle Cloud número de tu proyecto.

  3. Compila imágenes de Docker para los servicios embed-docs y chatbot. La imagen embed-docs contiene código de Python para la aplicación que recibe solicitudes del retransmisor de Eventarc y el trabajo de incorporación.

    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
    
  4. Verifica las imágenes:

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

    El resultado es similar a este:

    $REGION-docker.pkg.dev/$PROJECT_ID/${KUBERNETES_CLUSTER_PREFIX}-images/chatbot
    $REGION-docker.pkg.dev/$PROJECT_ID/${KUBERNETES_CLUSTER_PREFIX}-images/embed-docs
    
  5. Implementa una cuenta de servicio de Kubernetes con permisos para ejecutar trabajos de 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 -
    
  6. Cuando usas Terraform para crear el clúster de GKE y tienes create_service_account configurado como verdadero, se creará una cuenta de servicio independiente que usarán el clúster y los nodos. Otorga el rol artifactregistry.serviceAgent a esta cuenta de servicio de Compute Engine para permitir que los nodos extraigan imágenes del registro de Artifact Registry creado para embed-docs y chatbot.

    export CLUSTER_SERVICE_ACCOUNT=$(gcloud container clusters describe ${KUBERNETES_CLUSTER_PREFIX}-cluster \
    --location=${CONTROL_PLANE_LOCATION} \
    --format="value(nodeConfig.serviceAccount)")
    
    gcloud projects add-iam-policy-binding ${PROJECT_ID}  \
    --member="serviceAccount:${CLUSTER_SERVICE_ACCOUNT}" \
    --role="roles/artifactregistry.serviceAgent"
    

    Si no otorgas acceso a la cuenta de servicio, es posible que tus nodos tengan problemas de permisos cuando intenten extraer imágenes de Artifact Registry durante la implementación de los servicios embed-docs y chatbot.

  7. Implementa una implementación de Kubernetes para los servicios embed-docs y chatbot. Un Deployment es un objeto de la API de Kubernetes que te permite ejecutar varias réplicas de Pods que se distribuyen entre los nodos de un clúster:

    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 -
    
  8. Habilita los activadores de Eventarc para GKE:

    gcloud eventarc gke-destinations init
    

    Cuando se te solicite, ingresa y.

  9. Implementa el bucket de Cloud Storage y crea un activador de Eventarc con 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}
    

    Cuando se te solicite, escribe yes. Es posible que el comando tarde varios minutos en completarse.

    Terraform crea los siguientes recursos:

    • Un bucket de Cloud Storage para subir los documentos
    • Un activador de Eventarc
    • Una cuenta de servicio de Google Cloud llamadaservice_account_eventarc_namecon permiso para usar Eventarc.
    • Una cuenta de servicio de Google Cloud llamadaservice_account_bucket_namecon permiso para leer el bucket y acceder a los modelos de Vertex AI

    El resultado es similar a este:

    ... # Several lines of output omitted
    
    Apply complete! Resources: 15 added, 0 changed, 0 destroyed.
    
    ... # Several lines of output omitted
    

Carga documentos y ejecuta consultas de chatbot

Sube los documentos de demostración y ejecuta consultas para buscar en ellos con el chatbot:

  1. Sube el documento de ejemplo carbon-free-energy.pdf a tu bucket:

    gcloud storage cp vector-database/documents/carbon-free-energy.pdf gs://${PROJECT_ID}-${KUBERNETES_CLUSTER_PREFIX}-training-docs
    
  2. Verifica que el trabajo del incorporador de documentos se haya completado correctamente:

    kubectl get job -n ${DB_NAMESPACE}
    

    El resultado es similar a este:

    NAME                            COMPLETIONS   DURATION   AGE
    docs-embedder1716570453361446   1/1           32s        71s
    
  3. Obtén la dirección IP externa del balanceador de cargas:

    export EXTERNAL_IP=$(kubectl -n ${DB_NAMESPACE} get svc chatbot --output jsonpath='{.status.loadBalancer.ingress[0].ip}')
    echo http://${EXTERNAL_IP}:80
    
  4. Abre la dirección IP externa en tu navegador web:

    http://EXTERNAL_IP
    

    El chatbot responde con un mensaje similar al siguiente:

    How can I help you?
    
  5. Hacer preguntas sobre el contenido de los documentos subidos Si el chatbot no encuentra nada, responde I don't know. Por ejemplo, puedes hacer las siguientes preguntas:

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

    Un ejemplo de la respuesta del chatbot es similar al siguiente:

    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.
    
  6. Hazle al chatbot una pregunta que esté fuera del contexto del documento subido. Por ejemplo, podrías preguntar lo siguiente:

    You: What are Google plans to colonize Mars?
    

    Un ejemplo de la respuesta del chatbot es similar al siguiente:

    Bot: I don't know. The provided context does not mention anything about Google's plans to colonize Mars.
    

Acerca del código de la aplicación

En esta sección, se explica cómo funciona el código de la aplicación. Hay tres secuencias de comandos dentro de las imágenes de Docker:

  • endpoint.py: Recibe eventos de Eventarc en cada carga de documentos y, luego, inicia los trabajos de Kubernetes para procesarlos.
  • embedding-job.py: Descarga documentos del bucket, crea incorporaciones y las inserta en la base de datos de vectores.
  • chat.py: Ejecuta consultas sobre el contenido de los documentos almacenados.

En el diagrama, se muestra el proceso de generación de respuestas con los datos de los documentos:

En el diagrama, la aplicación carga un archivo PDF, lo divide en fragmentos y, luego, en vectores, y, por último, envía los vectores a una base de datos de vectores. Más tarde, un usuario le hace una pregunta al chatbot. La cadena de RAG usa la búsqueda semántica para buscar en la base de datos de vectores y, luego, devuelve el contexto junto con la pregunta al LLM. El LLM responde la pregunta y la almacena en el historial del chat.

Acerca de endpoint.py

Este archivo procesa mensajes de Eventarc, crea un trabajo de Kubernetes para incorporar el documento y acepta solicitudes desde cualquier lugar en el puerto 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)

Acerca de embedding-job.py

Este archivo procesa documentos y los envía a la base de datos de vectores.

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=os.getenv("VERTEX_AI_MODEL_NAME", "gemini-2.5-flash-preview-04-17"), 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("text-embedding-005")

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("text-embedding-005")

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("text-embedding-005")

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("text-embedding-005")

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)}")

Acerca de chat.py

Este archivo configura el modelo para que responda preguntas solo con el contexto proporcionado y las respuestas anteriores. Si el contexto o el historial de conversación no coinciden con ningún dato, el modelo devuelve 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=os.getenv("VERTEX_AI_MODEL_NAME", "gemini-2.5-flash-preview-04-17"), 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("text-embedding-005")

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=os.getenv("VERTEX_AI_MODEL_NAME", "gemini-2.5-flash-preview-04-17"), 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("text-embedding-005")

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=os.getenv("VERTEX_AI_MODEL_NAME", "gemini-2.5-flash-preview-04-17"), 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("text-embedding-005")

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})