Failover
The language model fails over between turns on a shared conversation context. Speech synthesis crosses to a second provider when you set a fallback voice.
The language model fails over between turns on a shared conversation context, so a switch does not lose the conversation. Speech synthesis crosses to a second provider for any agent with a fallback voice set. Synthesis region follows the voice and 1 of the 4 synthesis providers has an Australian endpoint, so a fallback voice can change where that part of a call runs. The component list is published on which parts of a call run in Sydney.
| Subsystem | Failover |
|---|---|
| Language model | Active, where the model has a second provider |
| Speech synthesis | Cross-provider, for agents with a fallback voice set |
Language model
Language model failover is active. It is built on a shared conversation context, so a switch does not lose the conversation.
Both paths are configured identically, including prompt, token limits, temperature and caching, so behaviour cannot drift between them.
A switch happens between turns and never in the middle of an utterance, so a caller does not hear a reply change character partway through.
If a provider stalls, a watchdog bounds the wait and switches. Once the provider recovers, traffic returns to the primary automatically.
Agents on a single provider
Some models have no second provider. For those agents, the first error causes the agent to speak a short closing line and end the call cleanly. The underlying error is recorded and never spoken.
Speech synthesis
Cross-provider voice failover is available. Set a fallback voice from a different provider on the Voice page. The fallback holds no connection until it is needed, so a healthy call consumes none of the fallback provider's capacity.
Failover switches once and stays, rather than returning the caller to a provider that was failing.