question for realtime models I usually inject a lot of exter LiveKit Community #ask-ai

question for realtime models - I usually inject a ...

flat-accountant-15519

12/23/2025, 6:30 AM

question for realtime models - I usually inject a lot of external information to the model via sandwiching "system messages" in the chat_ctx on conversation hooks like "on_user_turn_completed" etc. With live models - is it still possible to inject system messages to the chat_ctx and if so what hooks or events should i use given that "on_user_turn_completed" is not available.

tall-airline-16756

12/23/2025, 6:30 AM

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

tall-airline-16756

12/23/2025, 6:31 AM

Thanks — you're asking whether you can still inject system messages into

chat_ctx

for realtime models and which hooks to use since

on_user_turn_completed

isn't available. Short answer: yes — you can, but use different events. What to use • Use

on_input_speech_transcribed

(recommended). It fires after the realtime model produces a transcription and is the closest equivalent to

on_user_turn_completed

. Inject system messages there and call

await self.update_chat_ctx(chat_ctx)

. • Inject at session start with

on_enter()

for static context (e.g., user tier, persona). • You can also modify

chat_ctx

anytime before calling

generate_reply()

by copying, adding messages, then

update_chat_ctx()

. Why

on_user_turn_completed

isn't available •

on_user_turn_completed

requires client-side turn detection. Realtime models with server-side turn detection do not fire that hook, so use the transcription event instead. Small example (Python-style)

Copy code

async def on_input_speech_transcribed(self, transcription, event_ctx):
    ctx = self.chat_ctx.copy()
    ctx.add_message(role="system", content=f"[Note: user said: {transcription}]")
    await self.update_chat_ctx(ctx)

If you need, I can: • Provide a full example that includes RAG/DB lookup before injecting, • Show how to disable server-side turn detection to restore

on_user_turn_completed

flat-accountant-15519

12/23/2025, 6:32 AM

does gemini live native audio support this event? can you show an example in code how to do this?

tall-airline-16756

12/23/2025, 6:33 AM

Yes — Gemini Live native audio supports transcription events you can use to inject system messages. Use the session-level transcription event (

user_input_transcribed

on_input_speech_transcribed

) and update the chat context. Python example (AgentSession, final transcript):

Copy code

from livekit.agents import AgentSession, ChatMessage
from livekit.plugins import google

async def entrypoint(ctx):
    session = AgentSession(
        llm=google.realtime.RealtimeModel(
            model="gemini-2.5-flash-native-audio-preview-09-2025",
            voice="Puck",
            instructions="You are a helpful assistant."
        ),
    )

    @session.on("user_input_transcribed")
    async def on_user_transcribed(event):
        if event.is_final:
            system_msg = ChatMessage(role="system",
                                     content=f"[Note: user said: '{event.transcript}']")
            session.chat_ctx.add_message(system_msg)
            await session.update_chat_ctx(session.chat_ctx)

    await session.start(room=ctx.room)
    await ctx.connect()

Node.js/TypeScript example:

Copy code

import { AgentSession } from "livekit-agents";
import * as google from "@livekit/agents-plugin-google";

const session = new AgentSession({
  llm: new google.realtime.RealtimeModel({
    model: "gemini-2.5-flash-native-audio-preview-09-2025",
    voice: "Puck",
    instructions: "You are a helpful assistant."
  }),
});

session.on("user_input_transcribed", async (event) => {
  if (event.is_final) {
    session.chat_ctx.addMessage({
      role: "system",
      content: `[Note: user said: '${event.transcript}']`
    });
    await session.update_chat_ctx(session.chat_ctx);
  }
});

If you want, I can: • Provide a full RAG example that fetches data on transcription and injects it, or • Show how to disable server-side turn detection if you prefer client-side

on_user_turn_completed

flat-accountant-15519

12/23/2025, 6:33 AM

@wooden-afternoon-33967

tall-airline-16756

12/23/2025, 6:34 AM

Hi Narayan — I see your mention. Is there something specific you’d like me to add or clarify about the Gemini Live transcription example or event handling (code, RAG integration, or turn-detection options)?

5 Views

Open in Slack

Previous Next