Task 01: Create a Conversational RAG Flow

1. In Azure AI Studio, open the project created in Exercise 1 and select the </> Code option.

   

2. After selecting Code, you will create a compute instance to run VS Code in the cloud.

   

3. Select Authenticate.

   

4. After authentication, set up the VS Code container with configurations optimized for developing GenAI Apps.

   

5. Once set up, launch VS Code. In the example below, we start the Web version of VS Code, so you don’t need to have it installed on your local machine.

   

   If you prefer, you can also use VS Code on your desktop instead of the Web version.

1. After launching VS Code, clone the repository of your project created during the bootstrapping in Exercise 1.

2. Open the terminal in VS Code.

   

3. Execute the following commands:

   cd code
   git clone https://github.com/your_github_user/your_project_repo
   

   

   1) In AI Studio VS Code, store all your code in the code/ directory.

   2) Remember that your_github_user/your_project_repo was defined in the github_new_repo variable in the bootstrap.properties file from Exercise 1.

4. Your code is now loaded in VS Code. The src/chat_request.py file contains the Python program with the flex flow. You can review the get_response function to understand how the RAG flow is implemented.

   

1. Before starting development, load the data into the index in the development environment.

   We will load the files located in the data/sample-documents.csv directory of your project.

2. Open the terminal and perform the following steps:

   1. Update the Azure Developer CLI:

      curl -fsSL https://aka.ms/install-azd.sh | bash
      

   2. Log in to Azure CLI:

      az login --use-device-code
      

   3. With the –use-device-code option, navigate to https://microsoft.com/devicelogin in your browser and enter the code displayed in the terminal.

   4. Log in to Azure Developer CLI:

      azd auth login --use-device-code
      

      1. Similarly, visit https://microsoft.com/devicelogin to complete authentication.

   5. Add roles to your Service Principal

      principalId is the Enterprise Application Object ID

      rg='[your-your-resource-group-name]'
      principalId='[your-sp-objectId]'
      clientId='[your-sp-clientId]'
      clientSecret='[your-clientSequence]'
      tenantId='[your-tenantId]'
      subscriptionId='[your-subscriptionId]'
      
      # Service principal
      az login --service-principal --username $clientId --password $clientSecret --tenant $tenantId
      azd auth login --client-id $clientId --client-secret $clientSecret --tenant-id $tenantId
      
      scope="/subscriptions/$subscriptionId/resourceGroups/$rg"
      
      # Assign roles
      roles=(
      '2a2b9908-6ea1-4ae2-8e65-a410df84e7d1'  # Storage Blob Data Reader
      '8311e382-0749-4cb8-b61a-304f252e45ec'  # ACR Push Role
      '7f951dda-4ed3-4680-a7ca-43fe172d538d'  # ACR Pull Role
      '5e0bd9bd-7b93-4f28-af87-19fc36ad61bd'  # Cognitive Services OpenAI User
      'f6c7c914-8db3-469d-8ca1-694a8f32e121'  # Data Scientist
      'ea01e6af-a1c1-4350-9563-ad00f8c72ec5'  # Secrets Reader
      '8ebe5a00-799e-43f5-93ac-243d3dce84a7'  # Search Index Data Contributor
      '7ca78c08-252a-4471-8644-bb5ff32d4ba0'  # Search Service Contributor
      '64702f94-c441-49e6-a78b-ef80e0188fee'  # Azure AI Developer
      )
      
      for roleId in "${roles[@]}"; do
      az role assignment create \
         --assignee-object-id "$principalId" \
         --assignee-principal-type "ServicePrincipal" \
         --role "$roleId" \
         --scope "$scope"
      done
      

3. Navigate to the root of the project repo:

   cd your_project_repo
   

   From this point onward, all terminal commands will be executed within the code/your_project_repo directory, where your_project_repo is the name you chose for your project.

4. Initialize the environment variables with your development environment values:

   azd env refresh
   

   Ensure you use the same values for location, subscription, and environment name as used in the bootstrapping process.

5. Finally, execute the script to load the documents into AI Search:

   OpenAI and PromptFlow versions need to be updated in the requirement files. Open the file ./requirements.txt and update line 7 to: openai==1.56.2. Then, open the file ./src/requirements.txt and update lines 5 and 7 to: promptflow==1.17.1 and promptflow[azure]==1.17.1 respectively. Then run the pip install code to update the modules.

   pip install -r requirements.txt
   

   ./infra/hooks/postprovision.sh
   

   If you are trying to run this in your own computer instead of AI Studio’s VS Code, make sure to use 3.11.

1. Navigate to your repository directory and run:

   git checkout -b feature/feature_x
   

2. Open the src/chat.prompty file. This is the prompt for your RAG flow. Notice it is a generic prompt; you will create a specific prompt for your Lamna Health virtual assistant.

3. Replace the content of chat.prompty with the contents of the following file:

   updated_chat.prompty

   Notice that the new prompt provides better context for the assistant’s objectives.

1. Install the required libraries specified by our flow:

   pip install -r requirements.txt
   

2. Run the program with the flex flow:

   python src/chat_request.py
   

   

1. Set a breakpoint on the line where the flow is executed.

   

2. Start debugging.

   

3. While debugging, you can inspect variable contents, such as the documents retrieved during the AI Search retrieval process.

   

The first step is to use the @trace decorator in your function, as already done in the get_response function in the chat_request.py file.

1. Open chat_request.py and verify that the function is decorated with @trace.

   # chat_request.py
   
   @trace
   def get_response(question, chat_history):
       ...
   

   the ./src/flow.flex.yaml file has to add the following at the end due to a current bug

   environment_variables:
      PF_DISABLE_TRACING: false
   

2. Open the terminal and ensure you are logged into Azure.

    az login --use-device-code
   

3. Configure Prompt Flow to send trace data to your AI Project, remembering to replace the text in the brackets.

   pf config set trace.destination=azureml://subscriptions/[your_Subscription_id]/resourcegroups/[your_resource_group_name]/providers/Microsoft.MachineLearningServices/workspaces/[your_project_name]
   

4. Export the ./src directory to the PYTHONPATH to allow Python to find modules in the flow source directory.

   export PYTHONPATH=./src:$PYTHONPATH
   

   Skipping this step will result in a `ModuleNotFoundError: No module named chat_request’.

5. Enable Allow storage account key access option in Settings > Configuration of the storage account.

6. Execute the following command to run the flow with trace enabled. The run_flow.py script was created for ease of use.

   Before running the following command, ensure you have the subscription ID, resource group, and project name from your Azure AI Studio project exported in your shell environment.

      export AZURE_SUBSCRIPTION_ID=[your_subscription_id]
      export AZURE_RESOURCE_GROUP=[your_resource_group]
      export AZUREAI_PROJECT_NAME=[your_project_name]
   

       python ./util/run_flow.py "How can I access my medical records at Lamna Healthcare?"
   

   The output of the command will contain the link to the trace in AI Studio. you will need to grab it from there as there is no way to navigate to it directly from AI Studio.

   If you get a permission error, you may need to add the Storage Blob Data Contributor role to the user logged in with az login.

7. After running the flow, you can review the results in AI Studio.

   

8. Drill down into the trace for more detailed analysis.

   

9. Once done, you can revert the trace configuration to local.

   pf config set trace.destination="local"