What are all the pertinent events that are triggered before the agent response happens

Thanks for reaching out, Drew. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

how do these interact with preemptive generation

Thanks — I’ll explain how the events you asked about interact with preemptive generation. Summary • Preemptive generation starts LLM inference early using interim STT transcripts while the user is still speaking. • The same pre-response events still occur; preemptive output is either reused or discarded when the final turn completes. Step-by-step interaction 1. User starts speaking ->

user_state_changed

("speaking") and interim

user_input_transcribed

events emitted. These interim transcripts can trigger preemptive generation immediately if

preemptive_generation

is enabled. 2. STT streaming produces interim transcripts while preemptive LLM work runs in parallel. 3. Turn detection / endpointing determines the end of the user turn (VAD, STT endpointing, realtime model, or

session.commit_user_turn()

). 4. When a final transcript is produced ->

user_input_transcribed

with

is_final=true

. The system compares the final transcript to the interim one used for preemptive generation: ◦ If they match closely, the preemptive response is reused to reduce latency. ◦ If they differ significantly, the preemptive result is discarded and a fresh LLM inference is started with the final transcript. 1.

on_user_turn_completed

runs after final transcript and before agent reply. If this hook modifies context (RAG lookups, message edits, or cancels the reply via

StopResponse()

), any preemptive output is invalidated and a new response is generated reflecting the updated context. 2. Conversation item is added (

conversation_item_added

) and agent state changes to

thinking

. If preemptive output was valid it proceeds; otherwise a new LLM call runs, then TTS and playback. Interruption and cancellation • If the user interrupts mid-speech and interruption criteria are met (e.g.,

min_interruption_duration

), the pipeline cancels queued/ongoing speech and preemptive responses and restarts inference with the new input. Practical implications • Preemptive generation trades off latency vs. correctness: it speeds replies when interim transcripts are accurate, but is always verified against final transcript and context changes. • Any server-side changes in

on_user_turn_completed

will force regeneration to ensure accuracy. Would you like me to show where to enable/configure

preemptive_generation

and interruption thresholds in your SDK (JS or Python), or an example of handling

on_user_turn_completed

? Sources: examples/voice_agents | livekit/livekit_composite | DeepWiki | Agent speech and audio | LiveKit docs | Pipeline nodes and hooks | LiveKit docs

is there an event called agent_started_processing?

Good question — there is no event named

agent_started_processing

. Use these instead to detect when processing begins: •

agent_state_changed

with

new_state = "thinking"

— definitive signal that the agent has started LLM inference. •

conversation_item_added

— indicates the LLM has begun generating content (useful for detecting generation start). • Preemptive generation may cause an earlier transition to

thinking

(or earlier generation events) while the user is still speaking. Would you like a short SDK example (JS or Python) to listen for

agent_state_changed

and

conversation_item_added

? Sources: Agent session | LiveKit docs | Turns overview | LiveKit docs | examples/voice_agents | livekit/livekit_composite | DeepWiki | Agent speech and audio | LiveKit docs | Agent Builder | LiveKit docs