Azure Machine Learning and OpenAI

• You can select + Create in the top toolbar, or
• Select the + Create button in the center of the page under No Azure OpenAI to display.

1. From the Subscription dropdown list, select your Azure subscription.
2. Select an existing Resource group or select Create new to create one.
3. Enter a unique resource Name within the selected subscription. Only alphanumeric characters and hyphens are allowed. The value must be 2-64 characters long and cannot start or end with a hyphen.
4. Select the Pricing tier required for sizing and capacity. For more information, see View full pricing details.
5. Review the policy information and links from the Content review policy section.
6. Select the Next button to proceed.

1. From the Type section, choose the Selected networks, configure network security for your Azure AI services resource option to allow access from specific Virtual Network and Subnet.
2. Select Virtual Network from the list of existing virtual networks.
3. Select the subnet that you created for the application from the list of Subnets.
4. Under the Firewall section, you can leave the Address range field blank unless you wish to whitelist the source IP/CIDR.
5. Select the Next button to proceed.

1. Add Name and Value for tags. You can add multiple tags.
2. Select the Next button to proceed.

1. Enter a descriptive and unique Deployment name for the model.
2. Select the Deployment type as Standard to allow other configurations
3. Model version, AI resource and Content filter are set to default.
4. Set the Tokens per Minute Rate Limit for this deployment based on your requirements. For example, it is set 400K.
5. Enable dynamic quota.
6. Select Deploy to create the deployment with the selected settings.

1. Endpoint information, including the Target URI and Key, is required for further integration.

1. From the Subscription dropdown list, select your Azure subscription.
2. Select an existing Resource group or select Create new to create one.
3. Enter a Name for the workspace. The Name must be between 3 and 33 characters long. Its first character has to be alphanumeric, and the rest may contain hyphens and underscores. No whitespace is allowed.
4. Select Region to specify the physical location where the workspace is deployed.
5. Create a new Storage account or select an existing one in your subscription. A Storage account is used as the default datastore for the workspace.
6. Create a new Azure Key vault resource or select an existing one in your subscription. A Key vault stores sensitive information that the workspace needs.
7. Create a new Azure Application insights resource or select an existing one in your subscription. The workspace uses Azure Application insights to store monitoring information about your deployed models.
8. Optionally, create the Container registry resource or select an existing one in your subscription. A container registry stores Docker images that you use for training and deployments. To minimize costs, Azure creates a new Azure Container Registry resource only after you build your first image.
9. Select the Next : Inbound Access button to proceed.

1. Choose the All networks options as Public network access.
2. Select the Next : Outbound Access button to proceed.

1. Choose the Public options as Network isolation. With this option, Compute can access public resources and outbound data movement is unrestricted.
2. Select the Next : Encryption button to proceed.

1. Leave the Encrypt data using a customer managed key checkbox unselected.
2. Select the Next : Identity button to proceed.

1. Scroll down to Storage account access section, and then select the Credential-based access option as Storage account access type.
2. Leave the High business impact workspace checkbox unselected.
3. Select the Next : Tags button to proceed.

1. Add Name and Value for tags. You can add multiple tags.
2. Select the Next : Review + create button to proceed.

1. Enter an unique name in the Compute name field for your compute instance.
   1. It is a required field and the Compute name must be between 3 and 24 characters long.
   2. Valid characters are capital and lower case letters, digits, and hyphens.
   3. Compute name must start with a letter.
   4. Compute name must be unique across all existing computes within an Azure region.
   5. If a hyphen is used, it must be followed by at least one letter later in the name.
2. Choose CPU as Virtual machine type for CPU intensive operations.
3. Choose Standard_E4ds_v4 as Virtual machine size or choose a similar size that is suitable for data manipulation and training on medium to large datasets.
4. Select the Next button to proceed.

1. Enable the Auto shut down toggle after idle timeout.
2. Set Shutdown after to 60 Minutes.
3. Select the Next button to proceed.

1. The User assignment, Assigned identity and SSH are disabled by default.
2. Enable the Virtual network toggle, and then from the dropdown list, select your Virtual network and Subnet to configure this compute instance connection where access to Oracle Database is configured.
3. Enable the Root access and SSO.
4. Select the Next button to proceed.

1. Update the Connection value for Oracle Database connection string.

   Run the following code to install required Python libraries:

   %pip install -U oracledb %pip install datasets==3.6.0

   Run the following code to setup the environment with all required libraries:

   
   import json from datasets 
   import load_dataset
   import getpass
   import oracledb

Run the following code to test the database connectivity:

userpwd = getpass.getpass("Enter password: ")
connection = oracledb.connect(user="DEMO_USER", password=demo****password,
                              host="vm-*****-scan.ocidemoclients.ocidemovnet01.oraclevcn.com", port=1521, service_name="DEMOCDB1_DEMOPDB01.paas.oracle.com")
print("Successfully connected to Oracle Database")

Run the following code to load the McAuley-Lab/Amazon-Reviews-2023 meta_Clothing_Shoes_and_Jewelry data into Oracle Database using this approach:

dataset = load_dataset("McAuley-Lab/Amazon-Reviews-2023", "raw_meta_Amazon_Fashion", split="full", trust_remote_code=True)
# print(dataset[0])

userpwd = getpass.getpass("Enter password: ")
connection = oracledb.connect(user="DEMO_USER", password=demo****password,
                              host="vm-****-scan.ocidemoclients.ocidemovnet01.oraclevcn.com", port=1521, service_name="DEMOCDB1_DEMOPDB01.paas.oracle.com")
print("Successfully connected to Oracle Database")

cursor = connection.cursor()
insert_query = """
INSERT INTO product (main_category, title, average_rating, rating_number, features, description, price, images, videos, store, categories, details, parent_asin, bought_together) 
VALUES (:main_category, :title, :average_rating, :rating_number, :features, :description, :price, :images, :videos, :store, :categories, :details, :parent_asin, :bought_together)
"""

for record in dataset:   
    cursor.execute(insert_query, 
                    main_category=record['main_category'], 
                    title=record['title'], 
                    average_rating=record['average_rating'], 
                    rating_number=record['rating_number'],
                    features=json.dumps(record.get('features', [])), 
                    description=json.dumps(record.get('description', [])), 
                    price=record['price'], 
                    images=json.dumps(record.get('images', {})),
                    videos=json.dumps(record.get('videos', {})),
                    store=record['store'], 
                    categories=json.dumps(record.get('categories', [])), 
                    details=json.dumps(record.get('details', {})), 
                    parent_asin=record['parent_asin'], 
                    bought_together=record['bought_together']
    )


connection.commit()
cursor.close()
connection.close()

1. Update api_key and azure_endpoint values with the actual values from the Text embedding model details.
2. Update the Connection value for Oracle Database connection string.
   Note
   
   Generating vector embedding incurs compute costs based on data volume. Limit the number of records in Product table to reduce compute costs.
   Run the following code to install required libraries:

   
   %pip install openai --upgrade
   %pip install oracledb --upgrade

   Run the following code to setup Azure OpenAI and generate embedding:

   import pandas as pd
   import json
   import os
   import getpass
   import oracledb
   import openai
   import array
   import time
   
   # Set your OpenAI API Key
   api_type = "azure"
   api_key = "Adw3Y1QIG5ueN6TPUD7**************************w3AAABACOGiX0G"
   azure_endpoint = "https://demo-azure-openai-service.openai.azure.com/" 
   api_version = "2024-02-01"
   
   client = openai.AzureOpenAI(
           azure_endpoint=azure_endpoint,
           api_key=api_key,
           api_version=api_version
       )
   
   
   # Function to Generate the Embeddings
   deployment = "text-embedding-3-large"
   
   def generate_embeddings(text):
       response = client.embeddings.create(
           input=text,
           model=deployment
       )
       return response.data[0].embedding
   
   
   # Batch Size
   batch_size = 100
   record_count = 826108
   start_id = 824101
   end_id = 0
   
   # Load your dataset
   userpwd = getpass.getpass("Enter password: ")
   connection = oracledb.connect(user="DEMO_USER", password=demo****password,
                                 host="vm-*****-scan.ocidemoclients.ocidemovnet01.oraclevcn.com", port=1521, service_name="DEMOCDB1_DEMOPDB01.paas.oracle.com")
   print("Successfully connected to Oracle Database")
   
   cursor = connection.cursor()
   
   select_query = """select product_id, title from product where product_id between :startid and :endid"""
   
   
   for i in range(start_id, record_count, batch_size):
       start_id = i
       end_id = i + batch_size
   
       # Generate embeddings for each row in the "title" column
       binds = []
       for product_id, title in cursor.execute(select_query, startid=start_id, endid=end_id):
           response = generate_embeddings(title)
   
           vec = array.array("f", response)
           binds.append([vec, product_id])
   
       print("Embedding completed for " + str(end_id))
   
       connection.autocommit = True
   
       # Update the database with Vector Embeddings
   
       cursor.executemany(
           """update product
           set title_embedding = :1
           where product_id = :2""",
           binds,     
       )
   
       time.sleep(2)
   
   # Close the connection
   connection.commit()
   cursor.close()
   connection.close()

1. Update api_key and azure_endpoint values with the actual values from the Text embedding model details.
2. Update the Connection value for Oracle Database connection string.

   Run the following code to install required libraries:

   %pip install openai --upgrade %pip install oracledb --upgrade

   Run the following code to generate embedding on the input and perform the database search:

   import os
   import sys
   import getpass
   import array
   import oracledb
   import openai
   from skimage import io
   import matplotlib.pyplot as plt
   import requests
   
   
   # Set your OpenAI API Key
   api_type = "azure"
   api_key = "Adw3Y1QIG5ueN6TPUD7**************************w3AAABACOGiX0G"
   azure_endpoint = "https://demo-azure-openai-service.openai.azure.com/" 
   api_version = "2024-02-01"
   
   client = openai.AzureOpenAI(
           azure_endpoint=azure_endpoint,
           api_key=api_key,
           api_version=api_version
       )
   
   
   # Function to Generate the Embeddings
   deployment = "text-embedding-3-large"
   
   def generate_embeddings(text):
       response = client.embeddings.create(
           input=text,
           model=deployment
       )
       return response.data[0].embedding
   
   
   # Create a connection to the database
   userpwd = getpass.getpass("Enter password: ")
   connection = oracledb.connect(user="DEMO_USER", password=demo****password,
                                 host="vm-*****-scan.ocidemoclients.ocidemovnet01.oraclevcn.com", port=1521, service_name="DEMOCDB1_DEMOPDB01.paas.oracle.com")
   print("Successfully connected to Oracle Database")
   
   
   cursor = connection.cursor()
   
   search_query = """select product_id,
                     json_value(images, '$.hi_res[0]') as image_url,
                     title
                     from product
                     where json_value(images, '$.hi_res[0]') is not null
                     order by vector_distance(title_embedding, :1, EUCLIDEAN)                  
                     fetch first 5 rows only"""
   
   
   while True:
       # Get the input text to vectorize
   
       text = input("\Enter a phrase. Type quit to exit : ")
   
       if (text == "quit") or (text == "exit"):
           break
       
       if text == "":
           continue
       
       sentence = [text]
       response = generate_embeddings(sentence)
   
       vec = array.array("f", response)
   
       print("vector embedding generated")
   
       cursor.execute(search_query, [vec])
   
       result = cursor.fetchall()
   
       image_urls = []
       plt.rcParams["figure.figsize"] = [7.50, 3.50]
       plt.rcParams["figure.autolayout"] = True
   
       for record in result:
           image_url = record[1]
           urldata = requests.get(image_url).content
           print("Product Id: " + str(record[0]))
           a = io.imread(image_url)
           plt.imshow(a)
           plt.axis('off')
           plt.show()
   
       
   
   # Close the connection
   connection.commit()
   cursor.close()
   connection.close()

1. 01_Load_Data: Pull retail store catalog sample data from public data set and insert into Oracle Database.
2. 02_Generate_Vector_Embeeding: Generate vector data for product title and update it into product vector column.
3. 03_Semantic_Search: Generate vector embedding for user search text and select records from Product table with distance algorithm and display in output for matching result.