Decisions and Celeris

The tryAGI.OpenAI package supports the OpenAI Decisions API through client.Decisions.CreateDecisionAsync. It returns typed predicate, choice, score, and refusal answers, with token usage. The SDK also supports Celeris Decision through the same API.

OpenAI Decisions

using tryAGI.OpenAI;

using var client = new OpenAiClient("OPENAI_API_KEY");
var response = await client.Decisions.CreateDecisionAsync(
    model: "gpt-6-luna",
    input: "I was charged twice. Please refund one payment.",
    questions:
    [
        new QuestionParamPredicate
        {
            Name = "refund",
            Instructions = "The customer is asking for a refund.",
        },
    ]);

var answer = response.Answers[0];
if (answer.IsRefusal)
{
    Console.WriteLine($"Refused: {answer.PickRefusal().Name}");
}
else
{
    Console.WriteLine(answer.PickPredicate().Probability);
}

OpenAI currently documents Decisions as a public beta using gpt-6-luna. Use QuestionParamChoice with ChoiceOptionParam for fixed categories, or QuestionParamScore with ScoreLevelParam for ordered levels. Answers follow the order of the questions and echo their optional names. Choice values preserve string and boolean types. Scores are probability-weighted level indices, starting at zero; they can fall between levels.

Celeris Decision

Use CustomProviders.CelerisDecision and model celeris-1-decision. Its base URL differs from the Celeris chat endpoint. It calls /v1/decisions using the OpenAI format; it does not call the separate System One endpoint.

using System.Text.Json;
using tryAGI.OpenAI;

using var client = CustomProviders.CelerisDecision("CELERIS_API_KEY");
using var options = JsonDocument.Parse("""{"explain":true}""");
var response = await client.Decisions.CreateDecisionAsync(new DecisionRequest
{
    Model = "celeris-1-decision",
    Input = "I was charged twice. Please refund one payment.",
    Questions =
    [
        new QuestionParamPredicate
        {
            Name = "refund",
            Instructions = "The customer is asking for a refund.",
        },
    ],
    AdditionalProperties = new Dictionary<string, object>
    {
        ["x_celeris"] = options.RootElement.Clone(),
    },
});

Console.WriteLine(response.Answers[0].PickPredicate().Probability);
var extra = (JsonElement)response.AdditionalProperties["x_celeris"];
Console.WriteLine(extra.GetProperty("explanations")[0].GetString());

Use JsonElement for extension values to keep serialization compatible with the SDK's source-generated JSON context. Celeris returns one explanation per answer when x_celeris.explain is true; explanations add processing time.

For images, build DecisionInput from user messages with DecisionInputText and DecisionInputImage parts:

var input = new DecisionInput(null,
[
    new DecisionInputMessage
    {
        Content = new DecisionInputContent(null,
        [
            new DecisionInputText { Text = "The expense claim is for $13.50." },
            new DecisionInputImage
            {
                ImageUrl = "data:image/png;base64," + Convert.ToBase64String(imageBytes),
            },
        ]),
    },
]);

Celeris requires inline base64 data URLs. Set image resolution through x_celeris.max_soft_tokens; image detail does not control its resolution. Structured data should be serialized to a string for input. Celeris accepts up to 512 questions, or 64 with explanations, and up to 8 images. It does not support streaming, tools, audio, files, or message roles other than user. Unknown fields can cause validation errors, so omit chat-specific options. See the Celeris reference for current limits.

Celeris chat

Use the separate CustomProviders.Celeris factory for Celeris-1 chat completions and Responses:

using var client = CustomProviders.Celeris("CELERIS_API_KEY");
var response = await client.Chat.CreateChatCompletionAsync(new CreateChatCompletionRequestVariant2
{
    Model = "celeris-1",
    Messages = ["Rewrite as a search query: cheap flights Sydney to Tokyo in June."],
    MaxCompletionTokens = 256,
});

Set an explicit output budget for production calls; input plus requested output must fit the model's context window. Check Celeris's current parameter guidance for the model's budget limits. Chat can also use IChatClient; Decision calls use the dedicated Decisions surface. These factories do not imply that Celeris supports embeddings or every OpenAI endpoint. Both factories accept an optional caller-owned HttpClient, which remains usable after the SDK client is disposed.

Verification

The SDK's offline contract tests check both provider URLs, Bearer authentication, all three question types, boolean/string choice values, refusal parsing, inline image input, Celeris extension fields, and validation error preservation. They do not establish live provider acceptance, latency, accuracy, or billing.