Request a decision
model: the provider-prefixed model Relay selecteddecision_id: the identifier for this routing decision
Execute and correlate
1
Send request context
Provide the messages Relay needs to evaluate the workload. Do not include unrelated secrets or application state.
2
Validate the model
Confirm the returned model is allowed by your execution layer and maps to a configured provider.
3
Execute through your infrastructure
Call the selected provider through your application or existing gateway. Your infrastructure owns credentials, retries, timeouts, streaming, and provider errors.
4
Preserve the decision identifier
Attach
decision_id to the execution telemetry so Relay can connect the recommendation to the outcome.Failure behavior
Define an application-side timeout and deterministic fallback for cases where the decision service cannot be reached. Relay’s configured fallback informs routing decisions, but route-only does not take over your provider call.Configure candidate models
Control which models Relay may return and how it balances cost, quality, and latency.