Create my project and call the API

A project of your own, a model deployed in it, an endpoint: that is all it takes to call AI from your code. The project is created in the existing Foundry resource, never in a new resource — that would be a full-fledged Azure lab, with its own Resource Group, tags and declaration.

Create the project

1. Open the labs Foundry project, then create a project from the project selector. The suggested resource must be aif-azurefoundry-lab-szn: it holds the lab’s quotas and bill.

2. Give the project a recognizable name, your trigram followed by the topic: proj-qro-extraction-factures. The resource is shared, and nobody should have to ask who owns this project. Same spirit as lab naming.

3. Deploy a model from the model catalog. The deployment name is the one your code will use: keep the model name, without embellishing it.

4. Choose the deployment type knowingly: that is where cost and processing location are decided, see just below.

5. Copy the project endpoint, shown on the project home page. It looks like this:

   https://aif-azurefoundry-lab-szn.services.ai.azure.com/api/projects/<your-project>

The deployment type decides cost and location

The type	How it is billed	Where your prompts are processed
Global Standard	Per token, nothing without calls. The widest catalog and the lowest price.	Any geography where the model is deployed, so outside Switzerland and outside Europe.
Data zone Standard	Per token, nothing without calls.	Within the resource’s data zone, without leaving it.
Standard (regional)	Per token, nothing without calls.	Within the resource’s geography. The catalog is narrowest there.
Provisioned	By the hour, calls or not, as long as the deployment exists.	Depends on the chosen variant.
Batch	Per token, roughly half price, asynchronous.	Global processing.

For a test, Global Standard is the default choice: nothing to pay as long as you don’t make calls. A provisioned deployment is an hourly commitment, not to be created without prior approval from lab@qiminfo.ch.

“Global” means outside Switzerland

The resource is in Switzerland, which says nothing about where your prompts travel: a Global deployment has them processed wherever the model runs. With no client, personal or sensitive data in the prompts, this is not an issue — which is precisely why that rule applies here too. If the subject of your test is the processing location, pick a regional deployment and check the catalog.

Call it from your code

Authentication uses your Qim info account, via Azure CLI: no key to store, no key to leak.

1. Sign in. If your Azure CLI session is open on another tenant, add --tenant with the labs tenant.

   az login

2. Install the client and set the endpoint.

   pip install "azure-ai-projects>=2.3.0" azure-identity
   export AZURE_AI_PROJECT_ENDPOINT="https://aif-azurefoundry-lab-szn.services.ai.azure.com/api/projects/proj-qro-extraction-factures"

3. Call the model. The model parameter expects the name of your deployment, not that of the catalog model — they look alike, and it is the number one cause of 404s.

   import os
   
   from azure.ai.projects import AIProjectClient
   from azure.identity import DefaultAzureCredential
   
   project = AIProjectClient(
       endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
       credential=DefaultAzureCredential(),
   )
   
   client = project.get_openai_client()
   
   response = client.responses.create(
       model="gpt-5-mini",
       input="Summarize in three sentences what an Azure Resource Group is.",
   )
   
   print(response.output_text)

To check access without writing any code, a token and a request are enough.

TOKEN=$(az account get-access-token --scope https://ai.azure.com/.default --query accessToken -o tsv)

curl -X POST "$AZURE_AI_PROJECT_ENDPOINT/openai/v1/responses" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-5-mini", "input": "Hello"}'

The API key is a secret, the token isn’t

The portal also offers a key to call the resource. It grants access to the models to whoever holds it, and it carries nobody’s name: there is no way to tell who consumed what. If you do take one, it goes in an environment variable or a Git-ignored file, never in the code, never in a ticket, never in a message. The token obtained through az login expires on its own and stays tied to your account: prefer it.

The Foundry SDKs changed their API between versions 1.x and 2.x, and the examples floating around the web mix the two. Your project’s playground shows the code matching what is deployed, in the language of your choice: it is the only reference that never goes stale.

What goes wrong

The symptom	What is happening
The portal wants to create a new resource	You are in the full creation flow, not the project one. Start again from the aif-azurefoundry-lab-szn resource and create the project inside it.
404 on the model name	model expects the deployment name as it appears in your project, and the deployment must be in the project targeted by the endpoint.
401 despite az login	Your session is open on another tenant or another account: az account show will tell you. Run az login again with --tenant.
403 while the playground works	The portal and the SDK don’t go through the same permissions. Write to lab@qiminfo.ch with the project name.
The deployment fails for lack of quota	The resource’s tokens-per-minute quota is already allocated to existing deployments. Lower the requested quota, or ask lab@qiminfo.ch to arbitrate.
The model you want isn’t offered	It doesn’t exist in the resource’s region, or not with the chosen deployment type. A Global type often widens the catalog.
The bill climbs while nothing is running	A provisioned deployment, a hosted fine-tuned model, or a test loop left running. Check the project’s list of deployments before looking anywhere else.

When I’m done

Nothing expires on its own. Delete the deployments first, then the project: that is what stops billing and frees up quota for your colleagues.

What matters is kept elsewhere: your prompts, test sets and scripts belong in a Git repository. No backup is guaranteed in the lab, and a deleted project takes its conversations with it.

Applicable rules

No sensitive data

My responsibility

Never put real client data, personal data or any sensitive data in a lab.

MandatoryPR-GOVCLOUDLAB-A-01 / PR-GOVLABONPREM-A-01

Indicative budget

Recommendation

Indicative budget of about 200 CHF per month per lab.

Not blockingPR-GOVCLOUDLAB-A-01

Clean up and document

My responsibility

Stop or delete unused resources, and document the lab as you go.

Not blockingPR-GOVCLOUDLAB-A-01

A need beyond self-service — a missing model, a larger quota, a provisioned deployment, fine-tuning, a connection to internal data: write to lab@qiminfo.ch describing the test and how long it will last.