how can i get log & metrics for realtime agents

Thanks for reaching out, biberim45. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

how can i get agent standing time duration on room

To help you measure an agent's standing time duration in a room, here are some friendly methods you can try out: • Track Participant Lifecycle Events: You can use the SDK's

OnParticipantConnected

and

OnParticipantDisconnected

events to note down when the agent joins and leaves the room. For instance, in Go:

room := lksdk.ConnectToRoom(url, lksdk.ConnectInfo{...}, &lksdk.RoomCallback{
    OnParticipantConnected: func(rp *lksdk.RemoteParticipant) {
        if rp.Kind() == lksdk.ParticipantAgent {
            agentJoinedAt = time.Now()
        }
    },
    OnParticipantDisconnected: func(rp *lksdk.RemoteParticipant) {
        if rp.Kind() == lksdk.ParticipantAgent {
            agentLeftAt = time.Now()
            duration := agentLeftAt.Sub(agentJoinedAt)
            log.Printf("Agent standing time: %v", duration)
        }
    },
})

• Use Prometheus Metrics: You might want to enable Prometheus scraping to get session duration metrics like

session_duration_ms

which can be really helpful for observability. • Aggregate Metrics: The

UsageCollector

class is great for summarizing metrics and matching them with timestamps to find out session durations. • Cloud Billing: You can also check LiveKit Cloud's agent session minutes to confirm standing time, as it reflects active connection time. If you'd like, I’m here to help you with more details or guide you through implementing any of these methods—just let me know! Sources: cmd/lk/perf.go | pkg/telemetry/prometheus/rooms.go | Capturing metrics | LiveKit Docs | Capturing metrics | LiveKit Docs | Docs::Home

i now use to gemini live api and gpt realtime api how can i get metrics Usage: UsageSummary(llm_prompt_tokens=26418, llm_prompt_cached_tokens=0, llm_input_audio_tokens=171, llm_completion_tokens=443, llm_output_audio_tokens=0, tts_characters_count=0, tts_audio_duration=0.0, stt_audio_duration=0.0) i taked but output audio tokens i saw to 0 why

I understand it can be a bit confusing when

output_audio_tokens

shows as 0 in your

UsageSummary

. This usually happens because of how the

UsageCollector

gathers metrics. While you can find audio token details (like

outputTokenDetails.audioTokens

) in the raw

RealtimeModelMetrics

events from Gemini Live API and GPT Realtime API, sometimes the

UsageCollector

isn’t set up to include these in the final summary. Here are a couple of steps that might help: • Double-check that your

UsageCollector

is properly processing and adding up the

outputTokenDetails.audioTokens

from the metrics events. • If you’re working with the JavaScript version, you might want to extend the

UsageSummary

interface to add fields like

llmInputAudioTokens

and

llmOutputAudioTokens

, and update the

collect

method to include these values in the totals. Please don’t hesitate to reach out if you’d like some more detailed help with updating the

UsageCollector

or if you run into any other questions—I’m here to support you! Sources: plugins/google/src/beta/realtime/realtime_api.ts | livekit-plugins/livekit-plugins-google/livekit/plugins/google/beta/realtime...

how can i do check usagecollector for llm output audiotokens and how can i expand for collect this data

The

UsageCollector

in LiveKit is designed to help you effortlessly collect LLM output audio tokens when using

RealtimeModelMetrics

. To make sure you’re capturing this data smoothly, here’s what you can do: 1. Use a real-time LLM, like

openai.realtime.RealtimeModel()

, in your agent setup. 2. Attach the

UsageCollector

to the

metrics_collected

event of your agent session. 3. When you need it, simply retrieve the summary using

get_summary()

. Here’s a friendly example in Python to guide you:

usage_collector = metrics.UsageCollector()

@agent.on("metrics_collected")
def on_metrics(metrics: metrics.AgentMetrics):
    usage_collector.collect(metrics)

# Later, access the collected audio tokens
summary = usage_collector.get_summary()
print(f"Output audio tokens: {summary.llm_output_audio_tokens}")

If you have any questions or need a bit more help, please don’t hesitate to reach out—I’m here for you! Sources: livekit-plugins/livekit-plugins-aws/livekit/plugins/aws/experimental/realti...

Agent: Test, Room: Test_room_42172ea3-70f6-4d16-be8c-ec6fb87e073a, Usage: UsageSummary(llm_prompt_tokens=26418, llm_prompt_cached_tokens=0, llm_input_audio_tokens=171, llm_completion_tokens=443, llm_output_audio_tokens=0, tts_characters_count=0, tts_audio_duration=0.0, stt_audio_duration=0.0) this my get_summary result why i see llm_output_audio_tokens=0 actually agent talked to me but not increas llm_output_audio_tokens

Thanks for your question! The

llm_output_audio_tokens

field in

UsageSummary

shows 0 because it only counts audio tokens generated directly by the LLM itself. If your agent uses a typical setup where the LLM generates text and then a separate TTS system converts that text to speech, those audio tokens won’t appear in

llm_output_audio_tokens

. To see values greater than zero in

llm_output_audio_tokens

, please make sure that: • You’re using a real-time LLM model that natively produces audio, like with

RealtimeModel

configured for audio output. • The model is set up to output audio directly, instead of depending on a separate TTS system. If your agent relies on TTS, the spoken output won’t show in

llm_output_audio_tokens

, but you can track TTS usage through

tts_characters_count

. If anything is unclear or you’d like more help, just let me know—I’m here to support you! Sources: livekit-plugins/livekit-plugins-anthropic/livekit/plugins/anthropic/llm.py

i use realtimemodel and this is results for summary: Agent: Test, Room: Test_room_42172ea3-70f6-4d16-be8c-ec6fb87e073a, Usage: UsageSummary(llm_prompt_tokens=26418, llm_prompt_cached_tokens=0, llm_input_audio_tokens=171, llm_completion_tokens=443, llm_output_audio_tokens=0, tts_characters_count=0, tts_audio_duration=0.0, stt_audio_duration=0.0) as you see tts_character_count=0 and output_audio_tokens=0 how can i calculate ai usage counts

I’m happy to help you understand how AI usage counts are calculated when

llm_output_audio_tokens

and

tts_character_count

are zero. The

UsageCollector

still keeps track of other important metrics as usual. Here’s a friendly breakdown of how it works: • LLM Metrics: The system adds up

llm_prompt_tokens

,

llm_prompt_cached_tokens

, and

llm_completion_tokens

from

LLMMetrics

events, no matter what the audio token values are. • TTS Metrics: Even if

tts_character_count

is zero, the system might still record

tts_audio_duration

if audio was created in other ways. If no TTS processing happened, both values will simply stay at zero. • Zero Handling: The

UsageCollector

starts all fields in

UsageSummary

at zero and adds up metrics straightforwardly, so zero values are perfectly normal and don’t cause any issues. In your situation, seeing no

llm_output_audio_tokens

and

tts_character_count

likely means no audio was generated directly by the LLM or via TTS. But don’t worry—other metrics like

llm_prompt_tokens

and

llm_completion_tokens

still give a good picture of the LLM’s usage. If you have any more questions or want me to walk you through these metrics further, just let me know—I’m here to help! Sources: livekit-plugins/livekit-plugins-anthropic/livekit/plugins/anthropic/llm.py | Logs, metrics, and telemetry | LiveKit Docs