类class agentscope.realtime

GeminiRealtimeModel

class GeminiRealtimeModel(RealtimeModelBase)

源码Source agentscope/realtime/_gemini/_model.py:27–480

Gemini 实时模型，处理低延迟的实时语音与多模态交互。位于 agentscope.realtime 包，继承自 RealtimeModelBase。核心成员包括 connect、close、events、push_audio、push_text、push_tool_result。

Three protocol traits shape this adapter:

Nothing on the wire is addressable. An item id is minted here for each model turn and each user turn so every event carries one.
Generation is implicit: a finished user turn or a tool response starts the reply, so :meth:request_response sends nothing.
Interruption is the server's job — it reports serverContent.interrupted and drops the rest of the reply itself — so :meth:cancel_response is a no-op. With caller-owned turns the caller silences the reply locally instead, and the server is told not to interrupt on its own.

嵌套类摘要Nested Class Summary

类Class	类型Kind	说明Description
`Parameters`	类class	参数，处理低延迟的实时语音与多模态交互。 Tuneables surfaced to the UI.

构造器摘要Constructor Summary

构造器Constructor	说明Description
`def __init__(model: str, credential:` `GeminiCredential, parameters: 'GeminiRealtimeModel.Parameters \| None' = None, mode…`	初始化 `GeminiRealtimeModel` 实例。 Initialize the Gemini realtime model.

属性摘要Property Summary

属性Property	说明Description
`input_transcription_enabled`	读取 `input_transcription_enabled` 属性。 Whether Gemini input transcription is enabled.

属性摘要Attribute Summary

属性Attribute	类型Type	说明Description	值Value
`type`	—	类型字段。	`'gemini_realtime'`
`supports_text_input`	—	支持文本输入字段。	`True`

方法摘要Method Summary

方法Method	说明Description
`async def connect(instructions: str, tools: list[dict] \| None = None, **kwargs: Any) -> None`async	建立连接并完成初始化。 Open the WebSocket and send the setup message.
`async def close() -> None`async	关闭连接并释放相关资源。 Close the WebSocket and stop reading.
`async def events() -> AsyncIterator[me.ModelEvent]`async	返回事件。 Yield events until the session ends.
`async def push_audio(pcm: bytes) -> None`async	设置推送音频。 Send PCM16 mono audio at :attr:`input_sample_rate`. With caller-owned turns, the first chunk opens the activity window that :meth:`commit_turn` closes.
`async def push_text(text: str) -> None`async	设置推送文本。 Send a complete text turn, which the model answers at once.
`async def push_tool_result(block:` `ToolResultBlock) -> None`async	设置推送工具结果。 Send one ``functionResponse``; the model resumes on its own.
`async def commit_turn() -> None`async	返回提交轮次。 Close the activity window, which asks for the reply.
`async def request_response() -> None`async	返回请求响应。 No-op: a finished turn or a tool response is the request.
`async def cancel_response() -> None`async	返回取消响应。 No-op: the protocol has no cancel frame.

构造器详情Constructor Detail

GeminiRealtimeModel

def __init__(model: str, credential: GeminiCredential, parameters: 'GeminiRealtimeModel.Parameters | None' = None, model_card: RealtimeModelCard | None = None) -> None

初始化 GeminiRealtimeModel 实例。

Initialize the Gemini realtime model.

参数Parameters

名称Name	类型Type	说明Description
`model`	str	模型实例（类型 `str`）。The model name, e.g. `"gemini-3.1-flash-live-preview"`.
`credential`	GeminiCredential	凭证对象（类型 `GeminiCredential`）。The Gemini credential.
`parameters`默认default `None`	'GeminiRealtimeModel.Parameters \| None'	参数（类型 `'GeminiRealtimeModel.Parameters \| None'`）。Provider parameters; defaults are used if omitted.
`model_card`默认default `None`	RealtimeModelCard \| None	模型卡片（类型 `RealtimeModelCard \| None`）。The model card, looked up by name if omitted.

源码Source agentscope/realtime/_gemini/_model.py:76–108

属性详情Attribute Detail

type

 = 'gemini_realtime'

类型字段。

The type field.

源码Source agentscope/realtime/_gemini/_model.py:68

supports_text_input

 = True

支持文本输入字段。

The supports text input field.

源码Source agentscope/realtime/_gemini/_model.py:69

属性详情Property Detail

input_transcription_enabled

def input_transcription_enabled() -> bool

读取 input_transcription_enabled 属性。

Whether Gemini input transcription is enabled.

返回Returns bool

返回布尔值，表示判断结果。Returns a boolean result.

源码Source agentscope/realtime/_gemini/_model.py:72–74

方法详情Method Detail

connect

async def connect(instructions: str, tools: list[dict] | None = None, **kwargs: Any) -> None

async

建立连接并完成初始化。

Open the WebSocket and send the setup message.

参数Parameters

名称Name	类型Type	说明Description
`instructions`	str	指令（类型 `str`）。The instructions.
`tools`默认default `None`	list[dict] \| None	工具列表（类型 `list[dict] \| None`）。The tools.
`kwargs`	Any	关键字参数（类型 `Any`）。The kwargs.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:114–151

close

async def close() -> None

async

关闭连接并释放相关资源。

Close the WebSocket and stop reading.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:153–162

events

async def events() -> AsyncIterator[me.ModelEvent]

async

返回事件。

Yield events until the session ends.

返回Returns AsyncIterator[me.ModelEvent]

以异步流的形式返回me.ModelEvent。Returns a stream of me model event.

源码Source agentscope/realtime/_gemini/_model.py:164–170

push_audio

async def push_audio(pcm: bytes) -> None

async

设置推送音频。

Send PCM16 mono audio at :attr:input_sample_rate.

参数Parameters

名称Name	类型Type	说明Description
`pcm`	bytes	pcm（类型 `bytes`）。The pcm.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:176–192

push_text

async def push_text(text: str) -> None

async

设置推送文本。

Send a complete text turn, which the model answers at once.

参数Parameters

名称Name	类型Type	说明Description
`text`	str	文本（类型 `str`）。The text.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:194–203

push_tool_result

async def push_tool_result(block: ToolResultBlock) -> None

async

设置推送工具结果。

Send one functionResponse; the model resumes on its own.

参数Parameters

名称Name	类型Type	说明Description
`block`	ToolResultBlock	内容块（类型 `ToolResultBlock`）。The block.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:205–226

commit_turn

async def commit_turn() -> None

async

返回提交轮次。

Close the activity window, which asks for the reply.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:228–232

request_response

async def request_response() -> None

async

返回请求响应。

No-op: a finished turn or a tool response is the request.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:238–239

cancel_response

async def cancel_response() -> None

async

返回取消响应。

No-op: the protocol has no cancel frame.

返回Returns None

无返回值。No value is returned.

源码Source agentscope/realtime/_gemini/_model.py:241–242

嵌套类详情Nested Class Detail

Parameters

class Parameters

参数，处理低延迟的实时语音与多模态交互。位于 agentscope.realtime 包，继承自 RealtimeModelBase.Parameters。

Tuneables surfaced to the UI. Declared in agentscope.realtime. Extends RealtimeModelBase.Parameters.

继承成员Inherited Members

继承自Inherited from RealtimeModelBase

list_models, replay_history, input_sample_rate, output_sample_rate, supports_history_replay