Skip to main content

Select models

Boundary statement​

Catalog capability is descriptive; configured availability is executable. A model can appear in /api/models with context and tool metadata while its provider remains unconfigured. It becomes a routing candidate only when the provider and model are enabled and the required endpoint and credential resolve.

Addressing rules​

UAR accepts two model forms:

  • provider/model selects both sides explicitly, for example openai/gpt-5.4-mini.
  • model resolves only against the current default provider.

An unknown provider, unknown model, or malformed explicit address returns an OpenAI-compatible 404 model error. An omitted model resolves from the default provider's configured default model, then the global LLM model, then a suitable enabled model from that provider.

Packaged UI workflow​

  1. Configure the provider in Admin → Providers.
  2. Open Admin → Models at /admin/models.
  3. Read the configured section separately from the catalog section. Catalog rows marked not configured cannot establish live availability.
  4. Filter catalog metadata by provider or advertised capabilities. These filters help selection; they do not test the provider.
  5. Choose Add model, select a configured provider and a catalog model, and confirm. The model is appended to that provider's configuration.
  6. Choose default on a configured model when it should be the provider's default. Removing a model leaves it in the catalog so it can be enabled later.
  7. Select the provider/model in Chat and complete a real request before making an inference claim.

The comparison view compares catalog metadata. It does not benchmark models or prove that any listed provider is reachable.

API workflow​

QuestionRequestMeaning
What does the catalog know?GET /api/modelsCatalog models and descriptive capability data, including providers that may not be configured.
Which OpenAI-compatible model IDs are available?GET /v1/modelsModels belonging to providers currently configured and enabled.
What does one catalog model advertise?GET /v1/models/{model_id}Capability and limit metadata for a provider/model ID, or an OpenAI-compatible 404.
Which models are enabled on one provider?GET /api/uar/providers/{id}/modelsThe provider registry's model configuration.
How is a model enabled or made default?PUT /api/uar/providers/{id}Replace the provider configuration with the intended enabled models and default_model.

Use an explicit provider/model in requests when reproducible routing matters. Use a bare model only when the current default-provider dependency is intended.

Observable success and failure​

  • A configured model appears in its provider resource and can appear in /v1/models while that provider is enabled.
  • A default model is visible on the provider and in the Models UI.
  • A request for an unknown route fails with model_not_found; UAR does not silently substitute a different explicit provider/model.
  • Catalog capability data may be incomplete or stale relative to a provider's live service. Only an actual request observes execution support.

Realtime state and reload authority​

Model additions, removals, and default changes made through the packaged UI are provider API writes. They update the live provider registry and reconcile the UI from server state. With durable settings configured, the provider configuration survives reconstruction.

Catalog content is build/runtime discovery data, not operator state. Changing an external catalog source or startup configuration requires its owning refresh, reload, or process restart; it is not a live model-registration API.

Profile limits​

  • server-full includes the Models UI, provider APIs, and the full server product composition.
  • minimal includes the model/provider HTTP paths but not the admin UI as a profile claim.
  • embedded-mobile has no /v1/models or admin route. The host supplies the driver and provider/model metadata used by its embedded runtime.

No catalog count or server-profile result certifies another profile. Continue with Verify genuine inference.