- 📖 Full walkthrough: Decisions API in Azure OpenAI: A C# Walkthrough
- ⭐ If this saved you time, star the repo so other .NET developers can find it.
- For more C# Samples using other AI Models check out AI models for C# developers
- For my Free Ebook on Prompt Engineering check out Prompt Engineering for .NET Developers
A small .NET 10 console app that sends a support ticket to the Azure OpenAI Decisions API. It prints:
- the probability that the customer is asking for a refund (
refund_requested, a predicate) - the department that should handle the ticket (
billing,technical,shipping, orother) and its confidence - a severity score across three levels (Cosmetic, Workaround available, Fully blocked) and its confidence
If the model refuses to answer a question, the app says to send the ticket to a person instead of treating the refusal as a low probability.
git clone https://github.com/taswar/OpenAI-DecisionApi-CSharp-Sample.git
cd OpenAI-DecisionApi-CSharp-Sample
dotnet user-secrets set "AZURE_OPENAI_ENDPOINT" "https://<resource>.openai.azure.com"
dotnet runWithout an API key, the app signs in with Entra ID, so run az login first. See Configuration for all settings.
- .NET 10 SDK
- An Azure OpenAI resource with a
gpt-6-lunadeployment. GPT-6 Luna is currently the only model the Decisions API supports, and the API is generally available in Microsoft Foundry. - Either an API key for the resource, or Azure credentials that
DefaultAzureCredentialcan use (for example,az login) and permission to call the resource
The app reads these settings from user secrets or environment variables:
| Name | Required | Description |
|---|---|---|
AZURE_OPENAI_ENDPOINT |
Yes | Base URL of the Azure OpenAI resource, with no path, for example https://<resource>.openai.azure.com |
AZURE_OPENAI_DEPLOYMENT |
No | Deployment name. Defaults to gpt-6-luna |
AZURE_OPENAI_API_KEY |
No | API key. If not set, the app uses DefaultAzureCredential |
Set them with user secrets from the project folder:
dotnet user-secrets set "AZURE_OPENAI_ENDPOINT" "https://<resource>.openai.azure.com"
dotnet user-secrets set "AZURE_OPENAI_DEPLOYMENT" "<your-deployment-name>"
# Optional: leave this out to use DefaultAzureCredential instead
dotnet user-secrets set "AZURE_OPENAI_API_KEY" "<your-api-key>"dotnet runExample output:
refund_requested: 100% true
department: billing (confidence 100%)
severity: 0.98 (confidence 13%)
Program.cs POSTs to {AZURE_OPENAI_ENDPOINT}/openai/v1/decisions. If AZURE_OPENAI_API_KEY is set, the request uses an api-key header. If it isn't, the app gets a token for https://cognitiveservices.azure.com/.default and sends it as a bearer token. The body contains the deployment name (model), the ticket text (input), and three questions: refund_requested (predicate), department (choice), and severity (score). PrintAnswer reads each item in the response's answers array and handles it according to its type, which includes refusal.
The endpoint URL is wrong. The app appends /openai/v1/decisions to AZURE_OPENAI_ENDPOINT, so the value must be the base URL only.
- Wrong:
https://<resource>.openai.azure.com/openai(any path such as/openai,/openai/v1, or/openai/deployments/...is appended to the route and produces a bad URL) - Right:
https://<resource>.openai.azure.com
A trailing slash is fine because the app trims it. If the base URL is correct and you still get 404, check that AZURE_OPENAI_DEPLOYMENT matches the name of your deployment exactly.
Environment variables override user secrets. If AZURE_OPENAI_ENDPOINT is set in your shell, update or unset it there (echo $AZURE_OPENAI_ENDPOINT shows the current value), or the old value is still used.
To check which value the app sees, list the secrets:
dotnet user-secrets listThe request was sent without a Content-Length header. This is fixed in the current Program.cs, which sends the body with StringContent. If you see this error, make sure your copy includes that change and does not use JsonContent.
- If you use an API key, check that
AZURE_OPENAI_API_KEYbelongs to the same resource as the endpoint. - If you use keyless auth, run
az loginand check that the signed-in account has the Cognitive Services OpenAI User role on the resource. DefaultAzureCredentialuses the first credential that works, so a different cached login can be used. Sign out and sign in again with the right account if needed.
The response shape differs from what the parser expects. Print the raw response (or check the Unexpected answer: lines in the output) and compare it with PrintAnswer in Program.cs.
Looking for a different decision model, or a reranker for RAG? See the companion samples:
- Microsoft-Decision-1 C# Demo is a .NET 10 console app that uses Microsoft-Decision-1, the decision model in Microsoft Foundry, for fast decision scoring. It routes a support ticket to a team, scores its severity, and checks for repeat contact, using
DefaultAzureCredential. - Cohere Rerank v5 Pro & Fast C# Sample is a .NET 10 console app that reranks search results for RAG with Cohere Rerank v5 Pro and Fast on Microsoft Foundry, and uses a calibrated score to decide which passages reach the chat model.
Comparing this sample with the Decision-1 demo shows the same support-ticket triage done with the Azure OpenAI Decisions API (predicate, choice, and score questions) and with Microsoft-Decision-1 (choice, score, and noul questions) in C#.
This project is licensed under the MIT License. See LICENSE.