类class agentscope.embedding

EmbeddingModelBase

class EmbeddingModelBase(Generic[InputT])

源码Source agentscope/embedding/_embedding_base.py:34–457

向量化模型的基类。位于 agentscope.embedding 包，继承自 Generic[InputT]。核心成员包括 list_models。

Generic over :data:InputT so that text-only subclasses (EmbeddingModelBase[str]) and multimodal subclasses (EmbeddingModelBase[str | TextBlock | DataBlock]) expose the correct inputs type to the IDE.

Follows the same pattern as :class:~agentscope.model.ChatModelBase:

__call__ splits inputs into batches of size :attr:batch_size, calls :meth:_call_api for each batch concurrently via :func:asyncio.gather, and merges the results. Each batch call is wrapped with retry logic.
Subclasses only implement :meth:_call_api for a single batch — no batching or retry code needed.
Each subclass may override :meth:_get_retryable_exceptions to declare provider-specific retriable errors.

嵌套类摘要Nested Class Summary

类Class	类型Kind	说明Description
`Parameters`	类class	参数，与模型服务通信并返回生成或向量化结果。 Provider-specific tunables for embedding models.

构造器摘要Constructor Summary

构造器Constructor	说明Description
`def __init__(credential:` `CredentialBase, model: str, dimensions: int \| None, parameters: BaseModel \| None, context_size…`	初始化向量化模型基类。 Initialize the embedding model base class.

属性摘要Attribute Summary

属性Attribute	类型Type	说明Description	值Value
`credential`	CredentialBase	凭证字段。	—
`model`	str	模型字段。	—
`dimensions`	int	维度字段。	—
`context_size`	int	上下文大小字段。	—
`batch_size`	int	`batch_size` 字段。	—
`max_retries`	int	最大重试字段。	—
`retry_delay`	float	重试延迟字段。	—
`supports_multimodal`	bool	支持多模态字段。	`False`

方法摘要Method Summary

方法Method	说明Description
`async def __call__(inputs: list[InputT], **kwargs: Any) ->` `EmbeddingResponse`async	以可调用对象方式执行。 Embed a list of inputs with automatic batching and retry.
`def list_models(custom_yaml_dir: str \| None = None) -> list[EmbeddingModelCard]`	列出模型。 List candidate embedding models from YAML files.

构造器详情Constructor Detail

EmbeddingModelBase

def __init__(credential: CredentialBase, model: str, dimensions: int | None, parameters: BaseModel | None, context_size: int, batch_size: int, max_retries: int, retry_delay: float) -> None

初始化向量化模型基类。

Initialize the embedding model base class.

参数Parameters

名称Name	类型Type	说明Description
`credential`	CredentialBase	凭证对象（类型 `CredentialBase`）。The API credential used for authentication.
`model`	str	向量化模型的名称。The name of the embedding model.
`dimensions`	int \| None	向量维度（类型 `int \| None`）。The output embedding vector dimensions for this instance. Required and first-class — see the class docstring for the rationale of keeping `dimensions` outside :class:`Parameters`. For backward compatibility with older configs that stored `dimensions` inside :class:`Parameters`, `None` is accepted at the signature level and is back-filled from `parameters.dimensions` when present.
`parameters`	BaseModel \| None	参数（类型 `BaseModel \| None`）。Provider-specific non-dimensional parameters. When `None`, the default `Parameters()` is used.
`context_size`	int	上下文大小（类型 `int`）。Maximum input length (in tokens) per single input item.
`batch_size`	int	batch 大小（类型 `int`）。Maximum number of input items per API call. When `__call__` receives more items, it splits them into batches and calls :meth:`_call_api` concurrently.
`max_retries`	int	最大重试次数（类型 `int`）。The maximum number of retries for each batch API call. Only exceptions listed in :meth:`_get_retryable_exceptions` count against this budget.
`retry_delay`	float	重试延迟（类型 `float`）。Seconds to sleep between retry attempts.

源码Source agentscope/embedding/_embedding_base.py:104–182

属性详情Attribute Detail

credential

CredentialBase

凭证字段。声明为 CredentialBase。

The credential field.

源码Source agentscope/embedding/_embedding_base.py:72

model

str

模型字段。声明为 str。

The model field.

源码Source agentscope/embedding/_embedding_base.py:75

dimensions

int

维度字段。声明为 int。

The dimensions field.

源码Source agentscope/embedding/_embedding_base.py:78

context_size

int

上下文大小字段。声明为 int。

The context size field.

源码Source agentscope/embedding/_embedding_base.py:86

batch_size

int

batch_size 字段。声明为 int。

The batch size field.

源码Source agentscope/embedding/_embedding_base.py:89

max_retries

int

最大重试字段。声明为 int。

The max retries field.

源码Source agentscope/embedding/_embedding_base.py:92

retry_delay

float

重试延迟字段。声明为 float。

The retry delay field.

源码Source agentscope/embedding/_embedding_base.py:95

supports_multimodal

bool = False

支持多模态字段。声明为 bool。

The supports multimodal field.

源码Source agentscope/embedding/_embedding_base.py:98

方法详情Method Detail

call

async def __call__(inputs: list[InputT], **kwargs: Any) -> EmbeddingResponse

async

以可调用对象方式执行。

The inputs are split into chunks of :attr:batch_size.

参数Parameters

名称Name	类型Type	说明Description
`inputs`	list[InputT]	输入消息（类型 `list[InputT]`）。The input data to embed. For text-only models this is `list[str]`; for multimodal models it is `list[str \| TextBlock \| DataBlock]`. Any :class:`TextBlock` items are transparently unpacked to their `.text` field on entry, so subclasses' batching and `_call_api` only have to handle `str` (and `DataBlock` for multimodal variants).
`kwargs`	Any	关键字参数（类型 `Any`）。The kwargs.

返回Returns EmbeddingResponse

返回EmbeddingResponse。`EmbeddingResponse`:

源码Source agentscope/embedding/_embedding_base.py:197–260

list_models

def list_models(custom_yaml_dir: str | None = None) -> list[EmbeddingModelCard]

列出模型。

Each concrete subclass should live in its own provider subdirectory (e.g.

参数Parameters

名称Name	类型Type	说明Description
`custom_yaml_dir`默认default `None`	str \| None	自定义 yaml 目录（类型 `str \| None`）。Override the YAML directory.

返回Returns list[EmbeddingModelCard]

返回EmbeddingModelCard列表。`list[EmbeddingModelCard]`:

源码Source agentscope/embedding/_embedding_base.py:411–457

嵌套类详情Nested Class Detail

Parameters

class Parameters

参数，与模型服务通信并返回生成或向量化结果。位于 agentscope.embedding 包。

Intentionally empty in the base — dimensions is a contract property of an embedding model (its output vector size), not a tunable knob, so it lives directly on the instance via the required :paramref:__init__.dimensions argument. Subclasses extend this class to expose non-dimensional knobs (e.g. Gemini's task_type, Dashscope's text_type).

extra="allow" is set so old persisted configs (where dimensions lived in parameters) keep deserialising without raising; :meth:EmbeddingModelBase.__init__ extracts it back out for backward compatibility.