from __future__ import annotations
import logging
from typing import Any, Callable, Dict, List, Optional
from langchain_core._api.deprecation import deprecated
from langchain_core.callbacks import (
AsyncCallbackManagerForLLMRun,
CallbackManagerForLLMRun,
)
from langchain_core.language_models.llms import LLM
from langchain_core.load.serializable import Serializable
from langchain_core.pydantic_v1 import Extra, Field, SecretStr, root_validator
from langchain_core.utils import convert_to_secret_str, get_from_dict_or_env
from tenacity import (
before_sleep_log,
retry,
retry_if_exception_type,
stop_after_attempt,
wait_exponential,
)
from langchain_community.llms.utils import enforce_stop_tokens
logger = logging.getLogger(__name__)
def _create_retry_decorator(max_retries: int) -> Callable[[Any], Any]:
import cohere
# support v4 and v5
retry_conditions = (
retry_if_exception_type(cohere.error.CohereError)
if hasattr(cohere, "error")
else retry_if_exception_type(Exception)
)
min_seconds = 4
max_seconds = 10
# Wait 2^x * 1 second between each retry starting with
# 4 seconds, then up to 10 seconds, then 10 seconds afterwards
return retry(
reraise=True,
stop=stop_after_attempt(max_retries),
wait=wait_exponential(multiplier=1, min=min_seconds, max=max_seconds),
retry=retry_conditions,
before_sleep=before_sleep_log(logger, logging.WARNING),
)
[docs]def completion_with_retry(llm: Cohere, **kwargs: Any) -> Any:
"""使用tenacity来重试完成调用。"""
retry_decorator = _create_retry_decorator(llm.max_retries)
@retry_decorator
def _completion_with_retry(**kwargs: Any) -> Any:
return llm.client.generate(**kwargs)
return _completion_with_retry(**kwargs)
[docs]def acompletion_with_retry(llm: Cohere, **kwargs: Any) -> Any:
"""使用tenacity来重试完成调用。"""
retry_decorator = _create_retry_decorator(llm.max_retries)
@retry_decorator
async def _completion_with_retry(**kwargs: Any) -> Any:
return await llm.async_client.generate(**kwargs)
return _completion_with_retry(**kwargs)
[docs]@deprecated(
since="0.0.30", removal="0.3.0", alternative_import="langchain_cohere.BaseCohere"
)
class BaseCohere(Serializable):
"""用于Cohere模型的基类。"""
client: Any #: :meta private:
async_client: Any #: :meta private:
model: Optional[str] = Field(default=None)
"""要使用的模型名称。"""
temperature: float = 0.75
"""一个非负浮点数，用于调整生成过程中的随机程度。"""
cohere_api_key: Optional[SecretStr] = None
"""Cohere API密钥。如果未提供，将从环境变量中读取。"""
stop: Optional[List[str]] = None
streaming: bool = Field(default=False)
"""是否流式传输结果。"""
user_agent: str = "langchain"
"""用于发出请求的应用程序的标识符。"""
@root_validator()
def validate_environment(cls, values: Dict) -> Dict:
"""验证环境中是否存在API密钥和Python包。"""
try:
import cohere
except ImportError:
raise ImportError(
"Could not import cohere python package. "
"Please install it with `pip install cohere`."
)
else:
values["cohere_api_key"] = convert_to_secret_str(
get_from_dict_or_env(values, "cohere_api_key", "COHERE_API_KEY")
)
client_name = values["user_agent"]
values["client"] = cohere.Client(
api_key=values["cohere_api_key"].get_secret_value(),
client_name=client_name,
)
values["async_client"] = cohere.AsyncClient(
api_key=values["cohere_api_key"].get_secret_value(),
client_name=client_name,
)
return values
[docs]@deprecated(
since="0.1.14", removal="0.3.0", alternative_import="langchain_cohere.Cohere"
)
class Cohere(LLM, BaseCohere):
"""连接大型语言模型。
要使用，您应该安装``cohere`` python包，并设置环境变量``COHERE_API_KEY``为您的API密钥，或将其作为命名参数传递给构造函数。
示例:
.. code-block:: python
from langchain_community.llms import Cohere
cohere = Cohere(model="gptd-instruct-tft", cohere_api_key="my-api-key")
"""
max_tokens: int = 256
"""表示每代要预测的令牌数量。"""
k: int = 0
"""每个步骤考虑的最有可能的令牌数量。"""
p: int = 1
"""每一步需要考虑的标记的总概率质量。"""
frequency_penalty: float = 0.0
"""根据频率惩罚重复的标记。取值范围在0到1之间。"""
presence_penalty: float = 0.0
"""惩罚重复的标记。取值范围在0到1之间。"""
truncate: Optional[str] = None
"""指定客户端如何处理超过最大令牌长度的输入：从开头截断，结尾截断或不截断。"""
max_retries: int = 10
"""生成时最大的重试次数。"""
class Config:
"""此pydantic对象的配置。"""
extra = Extra.forbid
@property
def _default_params(self) -> Dict[str, Any]:
"""获取调用Cohere API的默认参数。"""
return {
"max_tokens": self.max_tokens,
"temperature": self.temperature,
"k": self.k,
"p": self.p,
"frequency_penalty": self.frequency_penalty,
"presence_penalty": self.presence_penalty,
"truncate": self.truncate,
}
@property
def lc_secrets(self) -> Dict[str, str]:
return {"cohere_api_key": "COHERE_API_KEY"}
@property
def _identifying_params(self) -> Dict[str, Any]:
"""获取识别参数。"""
return {**{"model": self.model}, **self._default_params}
@property
def _llm_type(self) -> str:
"""llm的返回类型。"""
return "cohere"
def _invocation_params(self, stop: Optional[List[str]], **kwargs: Any) -> dict:
params = self._default_params
if self.stop is not None and stop is not None:
raise ValueError("`stop` found in both the input and default params.")
elif self.stop is not None:
params["stop_sequences"] = self.stop
else:
params["stop_sequences"] = stop
return {**params, **kwargs}
def _process_response(self, response: Any, stop: Optional[List[str]]) -> str:
text = response.generations[0].text
# If stop tokens are provided, Cohere's endpoint returns them.
# In order to make this consistent with other endpoints, we strip them.
if stop:
text = enforce_stop_tokens(text, stop)
return text
def _call(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> str:
"""调用Cohere的生成端点。
参数:
prompt: 传递给模型的提示。
stop: 生成时可选的停止词列表。
返回:
模型生成的字符串。
示例:
.. code-block:: python
response = cohere("Tell me a joke.")
"""
params = self._invocation_params(stop, **kwargs)
response = completion_with_retry(
self, model=self.model, prompt=prompt, **params
)
_stop = params.get("stop_sequences")
return self._process_response(response, _stop)
async def _acall(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> str:
"""异步调用Cohere的生成端点。
参数：
prompt: 传递给模型的提示。
stop: 生成时可选的停止词列表。
返回：
模型生成的字符串。
示例：
.. code-block:: python
response = await cohere("Tell me a joke.")
"""
params = self._invocation_params(stop, **kwargs)
response = await acompletion_with_retry(
self, model=self.model, prompt=prompt, **params
)
_stop = params.get("stop_sequences")
return self._process_response(response, _stop)