Skip to main content

Model administration

The Models area controls how Starfire represents AI models as product resources rather than treating a provider model ID as the entire configuration.

Model record

A model can carry configuration such as:
  • provider
  • route
  • API model identifier
  • display name and description
  • context window
  • maximum output
  • tool support
  • vision support
  • reasoning support
  • streaming support
  • structured/JSON output
  • embeddings capability
  • research eligibility
  • FORGE eligibility

Pricing and metering

Model administration can also associate provider-cost and Starfire-metering metadata with the model. Examples include input/output cost data and a Starfire usage multiplier or equivalent pricing configuration. These values belong in administrative configuration rather than public documentation because provider prices and platform billing can change.

Availability

A model can be scoped to plan or rollout categories such as Free, Plus, Pro, Business, internal, or experimental access according to the active product configuration. Publishing a model does not mean every user receives it. Effective access still depends on plan, organization policy, user overrides, and feature rollout.

Routing configuration

Model configuration can reference a primary route, fallback route, timeout, retry policy, and circuit-breaker-style health behavior. Routing configuration should preserve required capabilities. A fallback that cannot satisfy tools, vision, context, or other requirements should not be considered equivalent merely because it produces text.

Health

Model health can include route/provider availability, latency, errors, rate limiting, and other operational signals. An unhealthy model can remain defined in the catalog while being unavailable for normal traffic.

Publishing changes

Model changes can affect user-visible availability, costs, and request behavior. Validate the model/provider relationship before broad publication. Useful checks include:
  • provider connectivity
  • model identifier correctness
  • streaming behavior
  • context/output limits
  • required capabilities
  • tool compatibility
  • metering behavior
  • fallback behavior
Do not place provider secrets into model descriptions or public documentation. Credentials belong in protected platform configuration.

Models & routing administration

Manage provider health, routes, fallbacks, and routing simulation concepts.