Gemini 3.8 Live 开发者指南

pip install --upgrade google-genai websockets numpy

import asyncio
from google import genai
from google.genai import types

PROJECT_ID = "PROJECT_ID"
LOCATION = "us-central1"
MODEL_ID = "gemini-3.8-live"

# Initialize the Gen AI client for enterprise.
client = genai.Client(enterprise=True, project=PROJECT_ID, location=LOCATION)

# Define a tool for asynchronous execution.
order_lookup_tool = types.FunctionDeclaration(
    name="lookup_order_status",
    description="Fetches real-time shipping status for a customer order ID.",
    parameters=types.Schema(
        type="OBJECT",
        properties={
            "order_id": types.Schema(
                type="STRING",
                description="The order ID, for example, ORD-8472",
            ),
        },
        required=["order_id"],
    ),
)

# Configure the live connection.
config = types.LiveConnectConfig(
    response_modalities=["AUDIO"],
    system_instruction=types.Content(
        parts=[
            types.Part.from_text(
                text=(
                    "You are an enterprise concierge. When you call a tool, "
                    "briefly acknowledge the request and tell the user what you're doing. "
                    "If a tool returns no results, inform the user before trying again. "
                    "Never issue more than two consecutive tool calls without speaking."
                )
            )
        ]
    ),
    speech_config=types.SpeechConfig(
        voice_config=types.VoiceConfig(
            prebuilt_voice_config=types.PrebuiltVoiceConfig(voice_name="Aoede")
        ),
    ),
    input_audio_transcription=types.AudioTranscriptionConfig(
        custom_vocabulary=["ORD-8472", "Gemini Live", "Express Shipping"]
    ),
    output_audio_transcription=types.AudioTranscriptionConfig(),
    tools=[types.Tool(function_declarations=[order_lookup_tool])],
)

async def handle_tool_call(session, function_call):
    """Executes backend calls asynchronously without blocking incoming audio."""
    order_id = function_call.args.get("order_id", "")
    await asyncio.sleep(2.0)  # Simulate backend API latency.

    if not order_id.startswith("ORD-"):
        result = {
            "status": "invalid_argument",
            "retryable": False,
            "message": f"Order ID '{order_id}' is invalid. Ask the user to verify.",
        }
    else:
        result = {
            "status": "ok",
            "retryable": False,
            "order_id": order_id,
            "shipping_status": "Out for delivery by 4:00 PM",
        }

    # Send tool response matching the call ID.
    await session.send_tool_response(
        function_responses=[
            types.FunctionResponse(
                id=function_call.id,
                name=function_call.name,
                response=result,
            )
        ]
    )

async def main():
    async with client.aio.live.connect(model=MODEL_ID, config=config) as session:
        print(f"Connected to {MODEL_ID}")

        # Send initial text or stream 16 kHz PCM audio chunks.
        await session.send_realtime_input(
            text="Hi! Can you check the status of order ORD-8472?"
        )

        async for message in session.receive():
            # Handle asynchronous tool calls.
            if message.tool_call:
                for call in message.tool_call.function_calls:
                    asyncio.create_task(handle_tool_call(session, call))

            # Process audio output chunks and transcriptions.
            if message.server_content:
                if (
                    message.server_content.output_transcription
                    and message.server_content.output_transcription.text
                ):
                    print(
                        f"Agent: {message.server_content.output_transcription.text}",
                        end="",
                        flush=True,
                    )

if __name__ == "__main__":
    asyncio.run(main())