千帆

Qianfan

继承自: CustomLLM

百度智能云千帆 LLM 平台支持的 LLM。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

class Qianfan(CustomLLM):
    """
    The LLM supported by Baidu Intelligent Cloud's QIANFAN LLM Platform.
    """

    access_key: str = Field(
        description="The Access Key obtained from the Security Authentication Center of Baidu Intelligent Cloud Console."
    )

    secret_key: str = Field(description="The Secret Key paired with the Access Key.")

    model_name: str = Field(description="The name of the model service.")

    endpoint_url: str = Field(description="The chat endpoint URL of the model service.")

    context_window: int = Field(
        default=DEFAULT_CONTEXT_WINDOW, description="The context window size."
    )

    llm_type: APIType = Field(default="chat", description="The LLM type.")

    _client = PrivateAttr()

    def __init__(
        self,
        access_key: str,
        secret_key: str,
        model_name: str,
        endpoint_url: str,
        context_window: int,
        llm_type: APIType = "chat",
    ) -> None:
        """
        Initialize a Qianfan LLM instance.

        :param access_key: The Access Key obtained from the Security Authentication Center
            of Baidu Intelligent Cloud Console.
        :param secret_key: The Secret Key paired with the Access Key.
        :param model_name: The name of the model service. For example: ERNIE-4.0-8K.
        :param endpoint_url: The chat endpoint URL of the model service.
            For example: https://aip.baidubce.com/rpc/2.0/ai_custom/v1/wenxinworkshop/chat/completions_pro .
        :param context_windows: The context window size. for example: 8192.
        :param llm_type: The LLM type. Currently, only the chat type is supported.
        """
        if llm_type != "chat":
            raise NotImplementedError("Only the chat type is supported.")

        super().__init__(
            model_name=model_name,
            endpoint_url=endpoint_url,
            context_window=context_window,
            access_key=access_key,
            secret_key=secret_key,
            llm_type=llm_type,
        )
        self._client = Client(access_key, secret_key)

    @classmethod
    def from_model_name(
        cls,
        access_key: str,
        secret_key: str,
        model_name: str,
        context_window: int,
    ):
        """
        Initialize a Qianfan LLM instance. Then query more parameters based on the model name.

        :param access_key: The Access Key obtained from the Security Authentication Center
            of Baidu Intelligent Cloud Console.
        :param secret_key: The Secret Key paired with the Access Key.
        :param model_name: The name of the model service. For example: ERNIE-4.0-8K.
        :param context_windows: The context window size. for example: 8192.
        """
        service_list = get_service_list(access_key, secret_key, ["chat"])
        try:
            service = next(
                service for service in service_list if service.name == model_name
            )
        except StopIteration:
            raise NameError(f"not found {model_name}")

        return cls(
            access_key=access_key,
            secret_key=secret_key,
            model_name=model_name,
            endpoint_url=service.url,
            context_window=context_window,
            llm_type=service.api_type,
        )

    @classmethod
    async def afrom_model_name(
        cls,
        access_key: str,
        secret_key: str,
        model_name: str,
        context_window: int,
    ):
        """
        Initialize a Qianfan LLM instance. Then asynchronously query more parameters based on the model name.

        :param access_key: The Access Key obtained from the Security Authentication Center of
            Baidu Intelligent Cloud Console.
        :param secret_key: The Secret Key paired with the Access Key.
        :param model_name: The name of the model service. For example: ERNIE-4.0-8K.
        :param context_windows: The context window size. for example: 8192.
            The LLMs developed by Baidu all carry context window size in their names.
        """
        service_list = await aget_service_list(access_key, secret_key, ["chat"])
        try:
            service = next(
                service for service in service_list if service.name == model_name
            )
        except StopIteration:
            raise NameError(f"not found {model_name}")

        return cls(
            access_key=access_key,
            secret_key=secret_key,
            model_name=model_name,
            endpoint_url=service.url,
            context_window=context_window,
            llm_type=service.api_type,
        )

    @classmethod
    def class_name(cls) -> str:
        """Get class name."""
        return "Qianfan_LLM"

    @property
    def metadata(self) -> LLMMetadata:
        """LLM metadata."""
        return LLMMetadata(
            context_window=self.context_window,
            is_chat_model=self.llm_type == "chat",
            model_name=self.model_name,
        )

    @llm_chat_callback()
    def chat(self, messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse:
        """
        Request a chat.

        :param messages: The chat message list. The last message is the current request,
            and the previous messages are the historical chat information. The number of
            members must be odd, and the role value of the odd-numbered messages must be
            "user", while the role value of the even-numbered messages must be "assistant".
        :return: The ChatResponse object.
        """
        request = build_chat_request(stream=False, messages=messages, **kwargs)
        resp_dict = self._client.post(self.endpoint_url, json=request.dict())
        return parse_chat_response(resp_dict)

    @llm_chat_callback()
    async def achat(
        self,
        messages: Sequence[ChatMessage],
        **kwargs: Any,
    ) -> ChatResponse:
        """
        Asynchronous request for a chat.

        :param messages: The chat message list. The last message is the current request,
            and the previous messages are the historical chat information. The number of
            members must be odd, and the role value of the odd-numbered messages must be
            "user", while the role value of the even-numbered messages must be "assistant".
        :return: The ChatResponse object.
        """
        request = build_chat_request(stream=False, messages=messages, **kwargs)
        resp_dict = await self._client.apost(self.endpoint_url, json=request.dict())
        return parse_chat_response(resp_dict)

    @llm_chat_callback()
    def stream_chat(
        self, messages: Sequence[ChatMessage], **kwargs: Any
    ) -> ChatResponseGen:
        """
        Request a chat, and the response is returned in a stream.

        :param messages: The chat message list. The last message is the current request,
            and the previous messages are the historical chat information. The number of
            members must be odd, and the role value of the odd-numbered messages must be
            "user", while the role value of the even-numbered messages must be "assistant".
        :return: A ChatResponseGen object, which is a generator of ChatResponse.
        """
        request = build_chat_request(stream=True, messages=messages, **kwargs)

        def gen():
            resp_dict_iter = self._client.post_reply_stream(
                self.endpoint_url, json=request.dict()
            )
            yield from parse_stream_chat_response(resp_dict_iter)

        return gen()

    @llm_chat_callback()
    async def astream_chat(
        self,
        messages: Sequence[ChatMessage],
        **kwargs: Any,
    ) -> ChatResponseAsyncGen:
        """
        Asynchronous request a chat, and the response is returned in a stream.

        :param messages: The chat message list. The last message is the current request,
            and the previous messages are the historical chat information. The number of
            members must be odd, and the role value of the odd-numbered messages must be
            "user", while the role value of the even-numbered messages must be "assistant".
        :return: A ChatResponseAsyncGen object, which is a asynchronous generator of ChatResponse.
        """
        request = build_chat_request(stream=True, messages=messages, **kwargs)

        async def gen():
            resp_dict_iter = self._client.apost_reply_stream(
                self.endpoint_url, json=request.dict()
            )
            async for part in aparse_stream_chat_response(resp_dict_iter):
                yield part

        return gen()

    @llm_completion_callback()
    def complete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponse:
        """
        Request to complete a message that begins with the specified prompt.
        The LLM developed by Baidu does not support the complete function.
        Here use a converter to convert the chat function to a complete function.

        :param prompt: The prompt message at the beginning of the completed content.
        :return: CompletionResponse.
        """
        complete_fn = chat_to_completion_decorator(self.chat)
        return complete_fn(prompt, **kwargs)

    @llm_completion_callback()
    async def acomplete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponse:
        """
        Asynchronous request to complete a message that begins with the specified prompt.
        The LLM developed by Baidu does not support the complete function.
        Here use a converter to convert the chat function to a complete function.

        :param prompt: The prompt message at the beginning of the completed content.
        :return: A CompletionResponse object.
        """
        complete_fn = achat_to_completion_decorator(self.achat)
        return await complete_fn(prompt, **kwargs)

    @llm_completion_callback()
    def stream_complete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponseGen:
        """
        Request to complete a message that begins with the specified prompt,
        and the response is returned in a stream.
        The LLM developed by Baidu does not support the complete function.
        Here use a converter to convert the chat function to a complete function.

        :param prompt: The prompt message at the beginning of the completed content.
        :return: A CompletionResponseGen object.
        """
        complete_fn = stream_chat_to_completion_decorator(self.stream_chat)
        return complete_fn(prompt, **kwargs)

    @llm_completion_callback()
    async def astream_complete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponseAsyncGen:
        """
        Asynchronous request to complete a message that begins with the specified prompt,
        and the response is returned in a stream.
        The LLM developed by Baidu does not support the complete function.
        Here use a converter to convert the chat function to a complete function.

        :param prompt: The prompt message at the beginning of the completed content.
        :return: A CompletionResponseAsyncGen object.
        """
        complete_fn = astream_chat_to_completion_decorator(self.astream_chat)
        return await complete_fn(prompt, **kwargs)

metadata property

metadata: LLMMetadata

LLM 元数据。

from_model_name classmethod

from_model_name(access_key: str, secret_key: str, model_name: str, context_window: int)

初始化一个 Qianfan LLM 实例。然后根据模型名称查询更多参数。

:param access_key: 从百度智能云控制台安全认证中心获取的 Access Key。 :param secret_key: 与 Access Key 配对的 Secret Key。 :param model_name: 模型服务的名称。例如: ERNIE-4.0-8K。 :param context_windows: 上下文窗口大小。例如: 8192。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@classmethod
def from_model_name(
    cls,
    access_key: str,
    secret_key: str,
    model_name: str,
    context_window: int,
):
    """
    Initialize a Qianfan LLM instance. Then query more parameters based on the model name.

    :param access_key: The Access Key obtained from the Security Authentication Center
        of Baidu Intelligent Cloud Console.
    :param secret_key: The Secret Key paired with the Access Key.
    :param model_name: The name of the model service. For example: ERNIE-4.0-8K.
    :param context_windows: The context window size. for example: 8192.
    """
    service_list = get_service_list(access_key, secret_key, ["chat"])
    try:
        service = next(
            service for service in service_list if service.name == model_name
        )
    except StopIteration:
        raise NameError(f"not found {model_name}")

    return cls(
        access_key=access_key,
        secret_key=secret_key,
        model_name=model_name,
        endpoint_url=service.url,
        context_window=context_window,
        llm_type=service.api_type,
    )

afrom_model_name async classmethod

afrom_model_name(access_key: str, secret_key: str, model_name: str, context_window: int)

初始化一个 Qianfan LLM 实例。然后根据模型名称异步查询更多参数。

:param access_key: 从百度智能云控制台安全认证中心获取的 Access Key。 :param secret_key: 与 Access Key 配对的 Secret Key。 :param model_name: 模型服务的名称。例如: ERNIE-4.0-8K。 :param context_windows: 上下文窗口大小。例如: 8192。百度开发的 LLM 都带有上下文窗口大小在其名称中。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@classmethod
async def afrom_model_name(
    cls,
    access_key: str,
    secret_key: str,
    model_name: str,
    context_window: int,
):
    """
    Initialize a Qianfan LLM instance. Then asynchronously query more parameters based on the model name.

    :param access_key: The Access Key obtained from the Security Authentication Center of
        Baidu Intelligent Cloud Console.
    :param secret_key: The Secret Key paired with the Access Key.
    :param model_name: The name of the model service. For example: ERNIE-4.0-8K.
    :param context_windows: The context window size. for example: 8192.
        The LLMs developed by Baidu all carry context window size in their names.
    """
    service_list = await aget_service_list(access_key, secret_key, ["chat"])
    try:
        service = next(
            service for service in service_list if service.name == model_name
        )
    except StopIteration:
        raise NameError(f"not found {model_name}")

    return cls(
        access_key=access_key,
        secret_key=secret_key,
        model_name=model_name,
        endpoint_url=service.url,
        context_window=context_window,
        llm_type=service.api_type,
    )

class_name classmethod

class_name() -> str

获取类名。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@classmethod
def class_name(cls) -> str:
    """Get class name."""
    return "Qianfan_LLM"

chat

chat(messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse

请求聊天。

:param messages: 聊天消息列表。最后一条消息是当前请求，之前的消息是历史聊天信息。成员数量必须为奇数，奇数消息的角色值必须为“user”，偶数消息的角色值必须为“assistant”。 :return: ChatResponse 对象。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_chat_callback()
def chat(self, messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse:
    """
    Request a chat.

    :param messages: The chat message list. The last message is the current request,
        and the previous messages are the historical chat information. The number of
        members must be odd, and the role value of the odd-numbered messages must be
        "user", while the role value of the even-numbered messages must be "assistant".
    :return: The ChatResponse object.
    """
    request = build_chat_request(stream=False, messages=messages, **kwargs)
    resp_dict = self._client.post(self.endpoint_url, json=request.dict())
    return parse_chat_response(resp_dict)

achat async

achat(messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse

异步请求聊天。

:param messages: 聊天消息列表。最后一条消息是当前请求，之前的消息是历史聊天信息。成员数量必须为奇数，奇数消息的角色值必须为“user”，偶数消息的角色值必须为“assistant”。 :return: ChatResponse 对象。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_chat_callback()
async def achat(
    self,
    messages: Sequence[ChatMessage],
    **kwargs: Any,
) -> ChatResponse:
    """
    Asynchronous request for a chat.

    :param messages: The chat message list. The last message is the current request,
        and the previous messages are the historical chat information. The number of
        members must be odd, and the role value of the odd-numbered messages must be
        "user", while the role value of the even-numbered messages must be "assistant".
    :return: The ChatResponse object.
    """
    request = build_chat_request(stream=False, messages=messages, **kwargs)
    resp_dict = await self._client.apost(self.endpoint_url, json=request.dict())
    return parse_chat_response(resp_dict)

stream_chat

stream_chat(messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponseGen

请求聊天，响应以流式返回。

:param messages: 聊天消息列表。最后一条消息是当前请求，之前的消息是历史聊天信息。成员数量必须为奇数，奇数消息的角色值必须为“user”，偶数消息的角色值必须为“assistant”。 :return: ChatResponseGen 对象，是 ChatResponse 的生成器。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_chat_callback()
def stream_chat(
    self, messages: Sequence[ChatMessage], **kwargs: Any
) -> ChatResponseGen:
    """
    Request a chat, and the response is returned in a stream.

    :param messages: The chat message list. The last message is the current request,
        and the previous messages are the historical chat information. The number of
        members must be odd, and the role value of the odd-numbered messages must be
        "user", while the role value of the even-numbered messages must be "assistant".
    :return: A ChatResponseGen object, which is a generator of ChatResponse.
    """
    request = build_chat_request(stream=True, messages=messages, **kwargs)

    def gen():
        resp_dict_iter = self._client.post_reply_stream(
            self.endpoint_url, json=request.dict()
        )
        yield from parse_stream_chat_response(resp_dict_iter)

    return gen()

astream_chat async

astream_chat(messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponseAsyncGen

异步请求聊天，响应以流式返回。

:param messages: 聊天消息列表。最后一条消息是当前请求，之前的消息是历史聊天信息。成员数量必须为奇数，奇数消息的角色值必须为“user”，偶数消息的角色值必须为“assistant”。 :return: ChatResponseAsyncGen 对象，是 ChatResponse 的异步生成器。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_chat_callback()
async def astream_chat(
    self,
    messages: Sequence[ChatMessage],
    **kwargs: Any,
) -> ChatResponseAsyncGen:
    """
    Asynchronous request a chat, and the response is returned in a stream.

    :param messages: The chat message list. The last message is the current request,
        and the previous messages are the historical chat information. The number of
        members must be odd, and the role value of the odd-numbered messages must be
        "user", while the role value of the even-numbered messages must be "assistant".
    :return: A ChatResponseAsyncGen object, which is a asynchronous generator of ChatResponse.
    """
    request = build_chat_request(stream=True, messages=messages, **kwargs)

    async def gen():
        resp_dict_iter = self._client.apost_reply_stream(
            self.endpoint_url, json=request.dict()
        )
        async for part in aparse_stream_chat_response(resp_dict_iter):
            yield part

    return gen()

complete

complete(prompt: str, formatted: bool = False, **kwargs: Any) -> CompletionResponse

请求补全以指定提示开始的消息。百度开发的 LLM 不支持 complete 函数。此处使用转换器将 chat 函数转换为 complete 函数。

:param prompt: 已完成内容开头的提示消息。 :return: CompletionResponse。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_completion_callback()
def complete(
    self, prompt: str, formatted: bool = False, **kwargs: Any
) -> CompletionResponse:
    """
    Request to complete a message that begins with the specified prompt.
    The LLM developed by Baidu does not support the complete function.
    Here use a converter to convert the chat function to a complete function.

    :param prompt: The prompt message at the beginning of the completed content.
    :return: CompletionResponse.
    """
    complete_fn = chat_to_completion_decorator(self.chat)
    return complete_fn(prompt, **kwargs)

acomplete async

acomplete(prompt: str, formatted: bool = False, **kwargs: Any) -> CompletionResponse

异步请求补全以指定提示开始的消息。百度开发的 LLM 不支持 complete 函数。此处使用转换器将 chat 函数转换为 complete 函数。

:param prompt: 已完成内容开头的提示消息。 :return: CompletionResponse 对象。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_completion_callback()
async def acomplete(
    self, prompt: str, formatted: bool = False, **kwargs: Any
) -> CompletionResponse:
    """
    Asynchronous request to complete a message that begins with the specified prompt.
    The LLM developed by Baidu does not support the complete function.
    Here use a converter to convert the chat function to a complete function.

    :param prompt: The prompt message at the beginning of the completed content.
    :return: A CompletionResponse object.
    """
    complete_fn = achat_to_completion_decorator(self.achat)
    return await complete_fn(prompt, **kwargs)

stream_complete

stream_complete(prompt: str, formatted: bool = False, **kwargs: Any) -> CompletionResponseGen

请求补全以指定提示开始的消息，响应以流式返回。百度开发的 LLM 不支持 complete 函数。此处使用转换器将 chat 函数转换为 complete 函数。

:param prompt: 已完成内容开头的提示消息。 :return: CompletionResponseGen 对象。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_completion_callback()
def stream_complete(
    self, prompt: str, formatted: bool = False, **kwargs: Any
) -> CompletionResponseGen:
    """
    Request to complete a message that begins with the specified prompt,
    and the response is returned in a stream.
    The LLM developed by Baidu does not support the complete function.
    Here use a converter to convert the chat function to a complete function.

    :param prompt: The prompt message at the beginning of the completed content.
    :return: A CompletionResponseGen object.
    """
    complete_fn = stream_chat_to_completion_decorator(self.stream_chat)
    return complete_fn(prompt, **kwargs)

astream_complete async

astream_complete(prompt: str, formatted: bool = False, **kwargs: Any) -> CompletionResponseAsyncGen

异步请求补全以指定提示开始的消息，响应以流式返回。百度开发的 LLM 不支持 complete 函数。此处使用转换器将 chat 函数转换为 complete 函数。

:param prompt: 已完成内容开头的提示消息。 :return: CompletionResponseAsyncGen 对象。

源代码位于 llama-index-integrations/llms/llama-index-llms-qianfan/llama_index/llms/qianfan/base.py

@llm_completion_callback()
async def astream_complete(
    self, prompt: str, formatted: bool = False, **kwargs: Any
) -> CompletionResponseAsyncGen:
    """
    Asynchronous request to complete a message that begins with the specified prompt,
    and the response is returned in a stream.
    The LLM developed by Baidu does not support the complete function.
    Here use a converter to convert the chat function to a complete function.

    :param prompt: The prompt message at the beginning of the completed content.
    :return: A CompletionResponseAsyncGen object.
    """
    complete_fn = astream_chat_to_completion_decorator(self.astream_chat)
    return await complete_fn(prompt, **kwargs)