One way I’ve been trying to measure latency (of the LLM) is by using the “console –text” and looking at the logs. But I believe that is not dumped into “Agent Insights”. Is this correct?

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

yes but that means I cannot do only text in an easy way no? if I do "uv run main.py dev" and connect via playground and send text, it will run STT-TTS

Thanks — good clarification. You're right to be thinking about how the dev/run modes interact with STT/TTS. Short answer: you can run text-only easily in dev mode and avoid STT/TTS. Running

uv run main.py dev

and sending text from the Playground does not trigger STT. Text sent from the frontend is published on the

lk.chat

text stream and handled as text. TTS only runs if audio output is enabled. What to do to measure LLM latency with text-only: 1. Disable audio input and output when you start the AgentSession so the agent neither listens for audio (no STT) nor synthesizes speech (no TTS). 2. Keep text input/output enabled so the Playground sends text and the agent replies over the

lk.transcription

text stream. Example configurations: • Python

await session.start(
    agent=MyAgent(),
    room=ctx.room,
    room_options=room_io.RoomOptions(
        audio_input=False,   # no STT
        audio_output=False,  # no TTS
        text_input=True,
        text_output=True,
    ),
)

• Node.js

await session.start({
    agent: new MyAgent(),
    room: ctx.room,
    inputOptions: { audioEnabled: false, textEnabled: true },
    outputOptions:{ audioEnabled: false, transcriptionEnabled: true },
});

How to receive agent text replies in the frontend: • Listen to the

lk.transcription

text stream (agent responses are published there when audio output is disabled). If you want latency measurements in Agent Insights: • Agent Insights only collects data from recorded sessions running in LiveKit Cloud with observability enabled. Local/dev

console --text

logs are not uploaded to Agent Insights. • If you need centralized latency metrics for text-only runs, either run the session as a recorded cloud session with observability enabled, or emit your own telemetry (timing events) to your metrics backend during the text-only flow. If you want, tell me: • Whether you want step-by-step changes for your current code (share the session start snippet), or • Whether you want to capture latency in Agent Insights (so I can explain how to run a recorded session and enable observability), or • Which telemetry backend you use so I can suggest how to emit timing events for local dev. Sources: Text and transcriptions | LiveKit Documentation | How to build an agent with speech-to-text input and text-only output | Live...

on another note. What kwargs can I pass to openai 5.3 ?

Thanks — I’m following up about your question on “OpenAI 5.3” kwargs. Quick clarification before I list parameters: by “OpenAI 5.3” do you mean • the OpenAI model ID (e.g., a hypothetical model named gpt-5.3), or • the OpenAI SDK/client version 5.3? I ask because the LiveKit docs don’t mention a model explicitly named “5.3.” Once you confirm which you mean, I can give accurate guidance. In the meantime, here’s what I can tell you that will likely be useful either way: • LiveKit lets you pass provider-specific parameters to LLMs via

extra_kwargs

on the

inference.LLM()

constructor (or equivalent provider config). Typical OpenAI parameters you can pass include: ◦

model

— the model ID to use. ◦

temperature

— controls randomness (common range 0–2). ◦

max_tokens

/

max_completion_tokens

— limit output length. ◦

top_p

,

presence_penalty

,

frequency_penalty

— sampling controls. ◦

stop

— stop sequences. ◦ Any provider-specific flags (for example:

reasoning_effort

,

tool_choice

,

parallel_tool_calls

) if supported by the integration. • If you want exact, model-specific kwargs (for a particular OpenAI model), confirm the model name and I’ll list the parameters you can pass through LiveKit and how to supply them (e.g.,

extra_kwargs={...}

). Which did you mean — model gpt-5.3 or SDK version 5.3? Also tell me whether you’re configuring the LLM in Python or Node, and I’ll give a concrete example. Sources: OpenAI LLM models | LiveKit Documentation