-
Egress/recording capacity is a hard limit; don’t make recording-start failure fatal to the call unless required.
-
Failover: parallel workers, health-check self-hosted, route new sessions to Cloud when saturated (in-flight may drop).
-
Colocated LiveKit Inference can deliver very low in-call TTFT for open models.
Connections
- uses Gemma 4
- related_to Large Language Model — LiveKit Inference can host models colocated with transport.
- related_to jambonz
- part_of Transport
- related_to Turn Taking — Built-in turn detector; compare vs Pipecat Smart Turn.