Is there a ebook of the whole Livekit documentation i can download?

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

when the agent session finishes how to get the tokens used ?

I understand you're looking to retrieve tokens used during an agent session, and I want to help clarify how this works with LiveKit. Currently, LiveKit doesn't offer a built-in feature to retrieve tokens after a session ends because tokens are usually short-lived and aren't stored for later access. However, a helpful approach is to log the tokens right when they're created through your token issuance service. This way, you'll have a handy record of all tokens issued for each session. If you'd like, I’d be more than happy to guide you through setting up this kind of logging—just let me know! Sources: Dispatching agents | LiveKit Docs | src/pages/api/token.ts | Analytics API | LiveKit Docs | livekit-plugins/livekit-plugins-anam/livekit/plugins/anam/api.py | complex-agents/nova-sonic/nova-sonic-form-agent/app/api/connection-details/...

no, I meant LLM tokens used, maybe session has it? for example session has history

To help you keep track of the LLM tokens used during your session, you can check out the token usage metrics through the

CompletionUsage

model. It includes handy fields like

prompt_tokens

,

completion_tokens

, and

total_tokens

. These metrics are generated during your session and can be captured in real-time. Also, you’ll find the entire conversation history stored in

session.history

. If you'd like, you can save this history to a file by using a shutdown callback once your session ends. If you’d like a hand setting this up or want some examples, just let me know—I’m here to help! Sources: livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/realtime/real... | Workflows | LiveKit Docs | Workflows | LiveKit Docs | Session recording and transcripts | LiveKit Docs

agent keeps self interrupting, how to fix it?

It sounds like you're dealing with an agent that keeps self-interrupting—let’s work through some settings that can help smooth things out: • Try setting

allow_interruptions=False

in your

Agent

or

AgentSession

configuration. This helps prevent interruptions during speech. • Using

resume_false_interruption=True

along with

false_interruption_timeout

can be really useful to manage false interruptions caused by VAD sensitivity. • If you’re using a Realtime LLM with built-in turn detection, it might conflict with your local interruption settings, so it’s good to avoid that combination. • Explicitly configuring turn detection, like setting

turn_detection="vad"

, gives you better control over how the conversation flows. If you’d like, I’m here to help you implement these changes or provide some examples—just let me know! Sources: Turn detection and interruptions | LiveKit Docs | livekit-agents/livekit/agents/voice/agent_activity.py

for some reason, the sound doesnt start playing for almost 500 ms using _background_audio_player for backround sound effects_ why ?

Thanks for reaching out! The 500ms delay you're noticing in sound playback when using _background_audio_player_ is most likely due to the default

min_endpointing_delay

setting in LiveKit's Voice Activity Detection (VAD) system. This setting creates a 500ms silence threshold before triggering actions like playing background sounds. If you'd like to reduce this delay, a great way is to lower the

min_endpointing_delay

by adjusting the

min_silence_duration

in your agent's turn detection settings. Here's an example to guide you:

agent = VoicePipelineAgent(
    vad=silero.VAD.load(min_silence_duration=300),
    stt=deepgram.STT(),
    llm=openai.LLM(),
    tts=openai.TTS(),
)

By setting this value to 300ms or even lower, your background sound effects will play more promptly after the user finishes speaking. If you have any questions or want a hand with this setup, please don't hesitate to ask—I’m here to help! Sources: Improving voice AI's turn detection with transformers | Improving voice AI's turn detection with transformers | VoicePipelineAgent | LiveKit Docs | Improving voice AI's turn detection with transformers | examples/drive-thru/drivethru_agent.py

this didnt help

Could you please help me understand a bit more about when the 500ms delay happens? For example: • Is it right after you call

play()

on the _background_audio_player_? • Does it occur after the user stops speaking? • Or maybe during continuous playback alongside other media? Knowing this will really help us figure out if the issue is related to system settings, resource limits, or something specific to the platform. Thanks so much for your patience! Sources: src/room/Room.ts | livekit-agents/livekit/agents/voice/background_audio.py | Improving voice AI's turn detection with transformers | Improving voice AI's turn detection with transformers | Improving voice AI's turn detection with transformers

How to add evals? What about evals for realtime models?

Thank you for reaching out! Could you help me understand a bit more about what you mean by "adding evals"? For example: • Are you thinking about evaluation metrics to check how well your model is performing? • Or maybe you're interested in adding some custom evaluation logic for your realtime models? • Or is it something specific within LiveKit that you're curious about? The more details you can share, the better I can assist you. I'm here to help!