First-call delay can begin before the conversation pipeline starts. Measure deployment startup, dispatch, process initialization, provider connection, and first audible output separately before changing turn settings.
LiveKit's server-options guide distinguishes process prewarming from model prewarming. Its deployment-management guide documents hosting cold-start behavior. Check current plan, environment, SDK version, and provider implementation before assuming a delay has one cause.
Define the timing boundary precisely
Record the caller or browser connection attempt, job acceptance, session readiness, greeting generation, and audible playback where observable. A low model response time can coexist with a slow worker startup or delayed phone route.
Use the same controlled scenario for cold and already-active conditions. Label those categories rather than mixing them into an average that hides the user's first experience. Do not publish improvement figures without actual measured results and conditions.
Work through a fictional slow first greeting
A staging deployment has been idle. The first controlled call waits noticeably before the agent joins, while the second starts more quickly. The team should inspect hosting and process timing before concluding that the prompt or speech provider is slow.
A different fixture joins quickly but waits during first model activity. That suggests another boundary to inspect. The proposed investigation records both paths and verifies which supported warm-up mechanism affects the observed stage.
| Timing fixture | Question to answer |
|---|---|
| First call after idle | Hosting or process startup involved? |
| Second consecutive call | What changed in readiness? |
| Agent joins quickly, silence remains | Model or output stage delayed? |
| Burst beyond warm capacity | New process initialization exposed? |
| Provider unavailable | Warm-up failure and fallback visible? |
Balance readiness against resource cost
Keeping prepared capacity can use memory and compute while reducing some initialization work. More warm processes do not automatically improve a bottleneck in a remote API or telephone path. Measure resource use and end-to-end timing for the actual model and workload.
Preloading client-specific content also needs isolation. Do not place one client's customer state in reusable shared process assets simply to avoid a lookup. Keep reusable assets separate from per-call authorised context.
Review workflow metrics to distinguish responsiveness from successful customer actions.
Questions about warm-up
Does prewarming remove all first-call delay?
No. It addresses supported initialization work, while hosting, routing, network, and provider stages may still contribute.
Should staging timing predict production timing?
Only after checking relevant environment differences. Current hosting behavior and capacity can differ.
What should the performance report contain?
Version, plan and environment, workload, cold or warm classification, stage boundaries, resource measurements, observed results, and unresolved uncertainty.