Themis.Inference.Ollama 4.0.2

Themis.Inference.Ollama

Use Themis.Inference.Ollama when an application needs typed System One decisions through an Ollama server, without the ONNX Runtime and tokenizer dependencies used by Themis.Inference.

System One decision inference

Use AddThemisOllamaSystemOneDecisionInference to register Ollama's typed decision endpoint. Register the shared decision router from Themis.Core and the provider directly; this path does not register ONNX runtime or model artifact services:

using Themis.Core;
using Themis.Core.Extensions;
using Themis.Inference.Ollama.Extensions;

services.AddThemisDecisionRouter(new DecisionRouterOptions("nimble")
{
    RouteUnspecifiedRequestsToDefaultModel = true
});
services.AddThemisOllamaSystemOneDecisionInference(
    options => options.Endpoint = new Uri("http://localhost:11434/"));

AddThemis remains available as the default ONNX convenience registration. For an existing application already using it, the provider can still be registered in its configureServices callback.

An explicit DecisionRequest.Model takes precedence over the configured router default (DecisionRouterOptions.DefaultModel in the standalone composition or ThemisOptions.DefaultModel through AddThemis). The endpoint requires Ollama 0.35.0 or later and a local model trained for System One. System One supports choice, score, and noul answers; choice criteria must contain 2-255 options (the model may impose a lower limit), and score criteria must contain 2-10 ordered levels. It does not support tools, streaming, or generation controls. A generic GGUF or chat model is not automatically a System One model; it must be supported by Ollama and trained/packaged for that decision API.

The adapter sends the original Themis state and namespaced extras in a versioned System One state envelope. The extras.rldc namespace includes the central question and attribute-driven prior results; application-defined namespaces are preserved. Task and Language remain routing hints and are not sent to the model. The System One model must be prepared to consume this envelope.

The adapter maps result probabilities to Themis answer-confidence semantics rather than reusing System One's distribution concentration value. System One input/output token usage is retained in DecisionUsage.