Model routing: route selection and configuration
Explains how Duale AI selects a Model Route from a Tenant's pool and links to route settings, structured output, and model change management.
This page covers how routing chooses a route, getting started with routing, key term definitions, and pre-enablement checks for Model Routes.
- How routing chooses a Model Route for a response.
- Getting started with task selection and route configuration.
- Definitions of models, routes, pools, and provider requests.
- Route declaration checks before production enablement.
Summaries were generated by AI.
For each model response a Task needs, Duale AI chooses one of the enabled Model Routes in the Tenant’s pool. Your
endpoint and API token select the deployment and the Tenant, and the routes enabled in that Tenant form the pool. A
RoutingPolicy expresses preferences inside that pool. It cannot add a route or guarantee one model. Duale AI’s own
analysis uses separate models outside your pool, as Scope and limits
describes.
How routing chooses a route
Duale AI keeps the routes that best fit the request, then orders them by your routing preferences, its own needs for that part of the Task’s work, benchmark evidence, measured speed, and configured price. When no route meets every preference, Duale AI uses the closest routes rather than failing. It never extends the Task deadline. Duale AI publishes no fixed weights or formula, and the order can change as the pool and the evidence change.
Start with the task you need to complete
- Choose a specialization: Task capabilities and benchmark evidence defines the work preferences and names the benchmarks associated with each.
- Submit a request: Integrate model routing provides a runnable SDK example and result checks. Contract and limits defines each preference.
- Configure a pool: Plan the deployment covers model selection and route setup. Use architecture patterns to choose Tenant or deployment boundaries.
- Configure a route: Route settings lists each setting that changes provider requests, cost estimates, or token counts.
- Shape a response schema: Structured output shows which provider types enforce your schema.
- Enable images: Images explains when an image reaches a model.
- Change a model: Control model changes covers evaluation, enablement, observation, reversal, retirement, and the change record.
Distinguish models, routes, pools, and requests
- Term
- Model
- Meaning
- A named inference capability offered by a provider; a family, release, mutable alias, and deployment can differ
- Term
- Model Route
- Meaning
- One configured path to a provider and model, with its endpoint, credential, prices, and declared capabilities
- Term
- Model pool
- Meaning
- The enabled Model Routes of one Tenant; every model turn of a Task uses only these routes
- Term
- Provider request
- Meaning
- One request that Duale AI sends to a selected Model Route. A model turn can make several provider requests, and one Task can use more than one route
A Model is a component used during a Task, not the Agent.
Check route declarations before enablement
Provider features describe what an endpoint can accept or produce; task capabilities describe the work you need. Saving a route validates its configuration, not its behavior, so qualify the route before you enable it in production.
Duale AI validates a completed answer against the response schema your application sent. A matching shape does not prove factual correctness, policy compliance, or business acceptance. Keep those checks in your application.
After enablement, observe the change with the signals available to your application, Duale AI, and the provider, and record who decided.