Skip to content

06 | Run Your Copilot

TODO: Validate these work with the latest CLI

Step 4: Run the co-pilot with a sample question

[!NOTE]
You can interact with the copilot using the Azure AI CLI (from commandline) or the Azure AI SDK (from Python code or Jupyter notebooks). Let's explore all three.

1. Using Azure AI CLI

The Azure AI CLI has an ai chat command that allows you to interact with the copilot. Once you have initialized resources and have a deployment, use ai chat to engage interactively or non-interactively with the copilot.

You can chat with your data or chat with a function for more grounded or custom chat experiences. Use ai chat help to see available options or browse the documentation for details.

Let's look at some examples.

  1. Example of a non-interactive chat. We'll use the --user option to ask the copilot a simple question.

    ai chat --user "What products do you have? Give me name, price and 1 sentence descriptions for each"

    Chat noninteractive

  2. Example of interactive chat. This sets up a conversation with the copilot that maintains context. Use exit to exit the chat.

    ai chat --interactive
    Chat noninteractive

  3. Example of a "chat with function" capability. This allows us to define and use a tailored function to interact with the copilot.

    ai chat --interactive --function src/copilot_aisdk/chat:chat_completion
    Chat completion

2. Using Azure AI SDK - Python Code

The src/copilot_aisdk/chat.py defines two functions using the Azure AI SDK, that effectively represent our copilot 'with custom data' implementation. - get_documents generates a vector embedding of the user question, then uses the Azure AI Search client to retrieve documents relevant to it. - chat_completion uses the Azure OpenAI chat deployment endpoint with the document-enhanced question, to get a more relevant res

The src/run.py file provides two capabilities: - It uses the Azure AI SDK to interact with Azure AI services (to create and manage resources) and with our copilot implementation (to ask user questions) - It has a main function that can be used to invoke these functions based on commandline arguments. Use src/run.py --help to see available options.

Let's ask our copilot a simple question using the --question option this provides, as follows:

python src/run.py --question "which tent is the most waterproof?"
Run py

Try adding a --stream option and see what happens.

(Hint: it will return a collection of objects, each with part of the response - allowing you to render that in chunks to simulate a more interactive chat experience for users.)

3. Using Azure AI SDK - Jupyter Notebook

We can also use the Azure AI SDK from Jupyter notebooks for a step-by-step learning experience that is ideal for exploration.

[!NOTE]
To execute code in a Jupyter notebook, we need a suitable kernel. The prebuilt dev container environment has a Python 3.10 kernel we'll use for this purpose, as shown below. For you own local environment follow this guidance to activate the right envorinment and kernel.

  1. We've provided a sample notebook in src/copilot_sdk/chat.ipynb. Start by opening that in VS Code then click the Select Kernel button at top right to get the drop-down menu, then select Python Environments.

    Open notebook

  2. Select the Python 3.10 environment (or higher) as your kernel in the drop-down

    Select Kernel 1. Check that the kernel is active (it should be shown in place of the Select Kernel button)

    Activate Kernel 1. Now you can execute the notebook cells one by one, using the Run Cell button at the top right of each cell. Let's start by executing the first cell which asks a simple question but has Stream set to true.

    Execute cell 1. The output cell will have an option to expand the output to see the full results. You should see something like this - showing how the response is streamed in chunks to allow us to render responses in an incremental, interactive way.

    Scroll results 1. Next, try executing the second cell which asks a simple question but has Stream set to false. This is the non-interactive chat mode, returning a single response as shown.

    Scroll results

You can now extend the notebook with more cells, each trying out copilot or Azure AI SDK functionality in an exploratory way.

Next Steps

🟩 Congratulations!

We've created a custom copilot based on our product data and validated its usage with a simple question from both Azure AI CLI and Azure AI SDK.

Now, we need to evaluate how it performs in comparison to other applications like chatgpt by using a relevant evaluation dataset.

Next Up: Evaluate Copilot Performance