ElBruno.AI.Decisions.Foundry 0.8.0

ElBruno.AI.Decisions.Foundry

Use Microsoft-Decision-1 on Microsoft Foundry or gpt-6-luna through the Azure OpenAI Decisions API with the shared .NET 10 IDecisionClient: Choose, Score and Assess. This is a community SDK, not an official Microsoft SDK.

dotnet add package ElBruno.AI.Decisions.Foundry --version 0.8.0

Quickstart

Deploy Microsoft-Decision-1 in Foundry and run az login. The endpoint-first constructor accepts either a Foundry project endpoint or resource root, along with the deployment name and an explicit token credential. Endpoint-first construction and project-endpoint support require version 0.8.0 or later:

using Azure.Identity;
using ElBruno.AI.Decisions.Foundry;

using var client = new FoundryDecisionClient(
    new Uri("https://<resource>.services.ai.azure.com/api/projects/<project>"),
    "<deployment-name>",
    new AzureCliCredential());
var teams = new Dictionary<string, string?>
{
    ["billing"] = "Invoices and payments",
    ["support"] = "Technical problems"
};
var result = await client.ChooseAsync(
    "Please correct my invoice.", "Which team should handle this request?", teams);
Console.WriteLine($"{result.Choice}: {result.Confidence:P1}");

URL template: https://<resource>.services.ai.azure.com/providers/microsoft/v1/systemone. Both the resource root (https://<resource>.services.ai.azure.com/) and project endpoint (https://<resource>.services.ai.azure.com/api/projects/<project>) resolve to this resource-scoped route, with or without a trailing slash. Full invocation URLs and custom paths are preserved. Deployment name: the deployment identifier sent to the API, for example msft-decision-1, not necessarily the catalog name Microsoft-Decision-1. Supply any Azure.Core.TokenCredential, including AzureCliCredential for local development or ManagedIdentityCredential in a hosted app. The constructor uses the credential you provide without selecting or replacing one.

Advanced configuration and custom transports

Use FoundryDecisionOptions for API keys, timeout configuration, or other advanced settings. A supplied API key takes precedence over FoundryDecisionOptions.Credential. AddFoundryDecisions uses the same resource/project endpoint resolution.

The HttpClient overload borrows the caller's client and does not dispose it. Keep it alive while requests run and dispose it afterward. The Foundry client sets its timeout to infinite and enforces FoundryDecisionOptions.Timeout per request. The ordinary options constructor creates and owns its transport.

Illustrative response

{
  "answers": {
    "decision": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 0.99, "support": 0.01 },
      "confidence": 0.95
    }
  }
}

Example output: billing: 95.0% (culture-dependent formatting). Confidence is distinct from the selected-label probability.

Choice, Score and Assess expose optional result.Usage from the top-level System One usage object. InputTokens, OutputTokens, TotalTokens, ReasoningTokens and CachedTokens remain null when unreported. Foundry may report input and output without a total; the SDK does not compute one. Reported zero is preserved. Availability varies by model and endpoint, and cross-model token counts do not imply equivalent computation or cost.

Tentative early access: Choice, Score and Assess passed a live API-key smoke test on October 9, 2026. Current source also passed live Azure CLI/Entra checks for Microsoft-Decision-1 and gpt-6-luna on October 26, 2026, including gpt-6-luna multi-question scoring and text/image input. These source checks do not validate a newly published package. This does not establish calibration or production readiness. The provider-neutral IDecisionClient methods expose one text question per call; use DecideAsync for image input or multi-question requests.

Complete sample | Configuration | Microsoft-Decision-1 | Repository

Azure OpenAI Decisions API

Available in 0.7.0: the APIs in this section are additive. Existing Microsoft-Decision-1 SystemOne callers keep their current configuration.

For gpt-6-luna, select the Azure OpenAI Decisions contract explicitly:

using var client = new FoundryDecisionClient(new FoundryDecisionOptions
{
    Api = FoundryDecisionApi.OpenAIDecisions,
    Endpoint = new Uri("https://<resource>.openai.azure.com"),
    Model = "<gpt-6-luna-deployment>"
});

The provider appends /openai/v1/decisions to a resource root. The default SystemOne mode remains for Microsoft-Decision-1. See Microsoft's Decisions API documentation.

For text/image input and multiple questions in one request, use the additive FoundryDecisionClient.DecideAsync API. It returns typed answers matched by question name, including per-question refusals and optional token usage. The provider-neutral IDecisionClient convenience methods remain text-only and send one question per request. See the complete Foundry sample.

For example, after creating the OpenAI client above:

FoundryDecisionResult result = await client.DecideAsync(
    FoundryDecisionInput.FromText("The package arrived with a cracked screen."),
    [
        new FoundryPredicateQuestion("damage", "Does the input describe product damage?"),
        new FoundryChoiceQuestion("team", "Which team should handle this?", [
            new FoundryDecisionChoice("support", "Product support"),
            new FoundryDecisionChoice("billing", "Payments and refunds")
        ])
    ]);

foreach (FoundryDecisionAnswer answer in result.Answers)
{
    if (answer is FoundryRefusalAnswer)
        Console.WriteLine($"{answer.Name}: refused, requires review");
    else if (answer is FoundryPredicateAnswer predicate)
        Console.WriteLine($"{predicate.Name}: {predicate.Probability:P1}");
    else if (answer is FoundryChoiceAnswer choice)
        Console.WriteLine($"{choice.Name}: {choice.Choice}");
}
if (result.Usage is { } usage)
    Console.WriteLine($"Reported total tokens: {usage.TotalTokens}");

Use 1-400 uniquely named questions. Results are returned in request order. Missing usage remains null. For images, use FoundryDecisionInput.FromMessages with FoundryDecisionMessage.User and FoundryDecisionContentPart.Text/Image. HTTPS image URLs must be publicly accessible; supported image data URLs are also accepted. The service allows up to 128 images and 32 MB combined image data. The SDK checks image count and URL shape, not decoding or combined byte size.

The walkthrough includes az login, separate Decisions:Foundry and Decisions:OpenAI settings, --openai, image input, and both offline modes.