ElBruno.AI.Decisions.Ollama 0.8.0

ElBruno.AI.Decisions.Ollama

Native local System One decisions through Ollama and the shared .NET 10 IDecisionClient API: Choose, Score and Assess. Uses /v1/systemone, not chat prompting or first-token logprobs.

dotnet add package ElBruno.AI.Decisions.Ollama --version 0.8.0
ollama pull nimble

Requires a running Ollama v0.35.0 or later and a local System One decision model such as nimble, Tev1, clef, or clef-flash, not a general-purpose chat model. No local API key is required.

Quickstart

using ElBruno.AI.Decisions.Ollama;

using var client = new OllamaDecisionClient(new OllamaDecisionOptions
{
    Endpoint = new Uri("http://localhost:11434/"),
    Model = "nimble"
});
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: http://<ollama-host>:11434/v1/systemone. Resource roots append that path; full invocation URLs are preserved.

Illustrative response (not a live test result)

{
  "model": "nimble",
  "answers": {
    "decision": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 0.9, "support": 0.1 },
      "confidence": 0.531
    }
  },
  "usage": { "input_tokens": 174, "output_tokens": 1 }
}

Example output: billing: 53.1% (culture-dependent formatting). Confidence measures distribution concentration, 1 - H(p) / ln(N), not calibrated correctness or the selected-label probability.

Choice, Score and Assess preserve top-level System One usage as optional result.Usage. The example reports InputTokens = 174, OutputTokens = 1 and TotalTokens = null, not 175. Missing counters, including reasoning and cached tokens, stay null; reported zero stays zero. Availability depends on the local model and server version. Token counts across models do not imply equivalent computation or cost.

Choice limits depend on the model (typically 2-26 options; API maximum 255). Score supports 2-10 ordered rubric levels and preserves fractional zero-based scores. Assess returns native Noul proposition probability without a hidden boolean threshold. Text requests are limited to 64 KiB, never truncated. The shared API exposes one text question per call, not images or batches.

Tentative early access: published 0.6.2 passed a live Choice, Score and Assess smoke test with nimble on Ollama 0.35.1 on October 13, 2026. Other models and server versions remain unverified. This does not establish calibration or production readiness. Offline contract checks do not prove model behavior. See the validation scope and results. Native System One support requires 0.6.1 or later; 0.6.0 used the old chat-logprob implementation.

Complete sample | Official decision guide | System One API | Repository