This repository contains miscellaneous script collections.

• /Azure Functions: This directory has Azure Function App projects.

  • /AzureFunctions/Proj_LLMSQL:

    • This Azure Function project contains Azure Function/s that articulate a response to natural language question on dataset within an Azure SQL Database tables. The function makes use of LangChain SQLDatabaseToolkit backed by gpt-35-turbo-16k (can work with 4k version too) LLM.
    • You require Azure OpenAI and Azure SQL Database resources provisioned beforehand.
    • Ensure the database contains tables (with descriptive names) that you would like to ask questions to.
    • This function can be called by any front-end application to render output.
    • To run this Azure Functions project locally
      • Install necessary prerequsites for Azure Functions
      • Create a local python virtual environment.
      • pip install -r requirements.txt
      • Rename local.settings_template.json file to local.settings.json and update the values in this file.
      • Open this repository in VS Code.
      • Open VS Code terminal and navigate to \miscellaneous\Azure Functions\Proj_LLMSQL directory
      • func start
      • Open Postman and create a new post request to the URL displayed in the terminal
        • Ensure that the question to the database is included in the request body
        • Request body example: {"question":"Which customers live in Paris?"}
        • Ensure to include function key as x-functions-key in the request header.
    • To deploy to Azure Function App
      • Create Azure Function App (Python 3.9)
      • Open VS Code terminal and navigate to \miscellaneous\Azure Functions\Proj_LLMSQL directory
      • az login
      • az account set --subscription [Your Subscription ID]
      • func azure functionapp publish [Your function app name]
      • Update application settings.

  • /AzureFunctions/GPT4V_Custom_Skill:

    • This Azure Function project contains function that accepts image as input (base64-encoded string format), passes it to GPT4-Vision endpoint and returns a text description based on the configuration parameters (system message, temperature, max response, etc.). The http trigger function is built to be called as Azure AI Search custom skill.
    • You require Azure OpenAI service with GPT4 Vision deployment, existing/new Azure AI Search service provisioned.
    • You will need to create additional AI Search artefacts to consume this functionality: Data source, Skillset, Indexer.
    • Please refer to information given on passing images to AI Search custom skills
    • Please review the public documentation to learn more about the GPT-4 Turbo with Vision
    • The provided Azure function code is tested for Python 3.10
    • Steps to run locally and deploy to Azure are same, listed in previous function section.

• /SynapseML: This directory contains Synapse ML project artefacts.

  • /SynapseML/ChatCompletion:

    • This is collection of artefacts to demonstrate the ChatCompletion using Azure OpenAI gpt-35-turbo (0613) on a spark dataframe in Synapse.

    • The code uses SynapseML library.

    • Setup instructions for SynapseML library and Spark MS SQL Connector have been given in the respective notebooks.

    • The sample dataset is a subset of OpinRank dataset (Hotel and car reviews) which is included in this repo.

    • You need to upload the provided dataset into your Synapse workspace (or a storage account connected to the workspace).

    • Ensure that Azure OpenAI service is provisioned and gpt-35-turbo (0613) model deployment is done beforehand.

    • Ensure you have a Linked Service created for an Azure Key Vault where the secrets have to be stored (AZSQLUSR, AZSQLPWD, OPENAISERVICENAME,OPENAIAPIKEY). Alternatively, you can provide them in code directly (not recommended).

    • Import both notebooks into your Synapse workspace.

      • First notebook imports provided dataset into an Azure SQL table.
      • Second notebook reads data from Azure SQL table, applies custom LLM prompt to each record of the dataframe's nominated column using ChatCompletion model.

    • Input dataset after import to Azure SQL server table, after successful run of the first notebook.
      (Query run in SQL Server Management Studio)

    • ChatCompletion System Message

      You are the AI analyst and lead the change management team in hospitality industry.
      

    • ChatCompletion Prompt

      Perform a thorough analysis of the comments written by our guests.
      
      If comment has a positive sentiment, identify what we need to be keep doing.
      If comment has a negative sentiment, identify what we need to improve.
      For neutral sentiment comment, say "N/A"
      
      Format your findings as per structure below:
      {
          "sentiment":"positive" / "negative" / "neutral",
          "identified_action": "",
          "action_type": "continue" / "improve" / "N/A"
      }
      
      Only return one finding per review comment.
      
      Provided review comments:
      

    • Output dataset created with SynapseML, saved into Azure SQL Server table after successful run of the second notebook. (Query run in SQL Server Management Studio)

    • The dataset / LLM prompt can be adjusted as per your requirement once you get this solution working.

    • The LLM prompt produces output in the JSON format. The T-SQL SELECT query to parse the output is also included in this repo.

• /AzureML/AutoML BatchEndpoints: This directory contains notebooks to train AutoML model and deploy it as a batch endpoint.

• /LargeDocumentQA: This directory contains demo python code to extract the text from small / large PDF documents, chunk and vectorize into AI Search Index for Document QA task.