Operate and Harden a Generative AI Solution in Azure

Table of Contents

1. Challenge

   Provision the Azure Function App that will host the pipeline code

   Create a Python Function App in the same region as the rest of the pipeline. Confirm the Function App's Application Insights instance is wired up for later observability. Scenario: The Function App is the compute layer that connects blob storage to the model and Cosmos DB. You'll create the host first, then deploy code into it.

2. Challenge

   Understand the critical-ticket logic added to the function code before deploying it

   Walk through the critical-ticket checks added to the pipeline and the branch that handles a ticket the model flags as critical. Learner will un-comment lines of code that focus on critical ticket flags for an interactive aspect. Trace the two new actions the pipeline takes for a critical ticket: skipping draft-response generation, and setting the is_critical flag on the document written to Cosmos DB. Scenario: This capstone extends the pipeline you already built. The base flow (blob trigger, model call, Pydantic shape, Cosmos upsert) is unchanged, so this review focuses only on what's new: the logic that detects a critical ticket and changes what the pipeline does with it.

3. Challenge

   Configure Azure Blob Storage as the ingestion point for raw support tickets.

   Decide on an appropriate redundancy and access tier for unpredictable, write-heavy ticket arrival. Create a dedicated container for inbound tickets and confirm it can receive blob uploads. Ensure that this resource and all other resources are deployed to the same region as the lab's Resource group. Scenario: You're setting up the landing zone for tickets that customer-facing systems will drop into storage. The container needs to be ready before any pipeline code exists.

4. Challenge

   Deploy the function app code from a local development environment

   Deploy the function code to the Function App and confirm the deployment completes successfully. Scenario: You'll deploy the same code you just reviewed. The supplied VM exists for learners who'd rather not configure a local Python and Functions Core Tools setup.

5. Challenge

   Provision Cosmos DB with the NoSQL API and confirm it's ready to receive pipeline writes

   Select the NoSQL API and explain why it fits the JSON document shape coming from a Pydantic-validated AI response. Select the Provisioned throughput capacity mode to comply with the lab environment's RU/s cap requirement. Create the database and container inside of the Cosmos DB account. Scenario: Your pipeline writes one document per processed ticket, with traffic that's bursty and hard to forecast. You'll select the capacity mode that fits both that pattern and the lab's cost constraints.

6. Challenge

   Provision the Microsoft Foundry resource and deploy a gpt-5-mini model to it

   Create the Microsoft Foundry resource in the same region as your existing storage and Cosmos DB resources. Deploy the gpt-5-mini model and capture the endpoint and deployment name your function code will need. Scenario: This is the AI layer of the pipeline. The endpoint and deployment name you create here will feed directly into the function app's environment variables in a later step.

7. Challenge

   Configure the Function App's environment variables to connect every component of the pipeline

   Supply the storage connection string, the Microsoft Foundry endpoint and deployment name, the Cosmos DB connection details, and the Application Insights connection string. Confirm each value matches the resource it points to. Scenario: Every service the function talks to is referenced through an environment variable. Getting these right is what turns a collection of resources into a working pipeline.

8. Challenge

   Run the pipeline end to end and trace a ticket through every stage

   Upload a sample email text file to the inbound blob container. Watch the function execute in real time using the Function App's log stream, then review the invocation history. Use Application Insights to inspect the end-to-end trace for the run. Open the Cosmos DB container and confirm the structured document the model produced has landed correctly. Upload a critical email ticket and verify the pipeline successfully flags the email as critical. Scenario: This is the payoff. A single file upload triggers the full pipeline, and you'll follow it through every layer — function logs, Application Insights, and finally the document that lands in Cosmos DB.