Switchyard and policies
Prepare intelligent routing
switchyard is a reserved public name. Choose a compatible classifier and enable approved catalog targets. Verify each target's capabilities, protocol, maximum estimated cost and reasoning levels.
The judge assesses candidate eligibility. Selection applies capability and scope constraints, then compares reasoning effort and the same maximum estimates used for credit reservations. It does not promise the lowest final cost.
Reasoning effort
Portable levels are none, minimal, low, medium, high, xhigh and max. Configure native parameters actually supported by each target. The gateway does not invent a universal equivalence between thinking budgets and portable levels.
In APIs, reasoning_effort takes precedence over the supported alternative effort fields. A request with only a thinking budget must also identify portable effort when routing requires it.
Cost and context
Classification and answer requests both pass through PII checks and credit reservations. Classifier cost reduces the request's remaining budget. Uncertain usage stops the flow and retains its reservation for reconciliation.
Linked usage records distinguish routing and answer purpose and share a request group. See Streaming and tools for diagnostic headers.
A Responses continuation or signed artifact may be route-bound. Do not transparently move it to an incompatible model. Revoking the target may prevent further continuations.
Configure policies
In Modelli e policy, choose a scope and resource allowed to your role. Set authorized models, default, PII protection and maximum request cost.
| Choice | Effect |
|---|---|
| Explicit public-name list | Allows selected names within higher-level restrictions |
| All allowed models | Includes future models, still within scope and higher policies |
| Default model | Establishes the expected choice when a client does not provide another allowed model |
| Per-request maximum | Limits reservation exposure |
| Required PII | Enforces applicable protection rules |
Policies support Netter, reseller, company, group and user scopes. Legacy API fields profile, allowed_profiles and default_profile now contain free public names rather than fixed capability tiers.
Embeddings and model changes
Keep embeddings bindings consistent with the indexes using them. Changing a model or namespace may require dataset reindexing; complete the authorized reindexing flow before using a new binding. Switchyard does not automatically replace named embeddings or audio models.