Part1

LangChain官网的解释

Part2

两者的选择

Part3

LangChain中ChatOpeAI和OpenAI支持的模型

@staticmethod
    def modelname_to_contextsize(modelname: str) -> int:
        ...
        model_token_mapping = {
            "gpt-4": 8192,
            "gpt-4-0314": 8192,
            "gpt-4-0613": 8192,
            "gpt-4-32k": 32768,
            "gpt-4-32k-0314": 32768,
            "gpt-4-32k-0613": 32768,
            "gpt-3.5-turbo": 4096,
            "gpt-3.5-turbo-0301": 4096,
            "gpt-3.5-turbo-0613": 4096,
            "gpt-3.5-turbo-16k": 16385,
            "gpt-3.5-turbo-16k-0613": 16385,
            "gpt-3.5-turbo-instruct": 4096,
            "text-ada-001": 2049,
            "ada": 2049,
            "text-babbage-001": 2040,
            "babbage": 2049,
            "text-curie-001": 2049,
            "curie": 2049,
            "davinci": 2049,
            "text-davinci-003": 4097,
            "text-davinci-002": 4097,
            "code-davinci-002": 8001,
            "code-davinci-001": 8001,
            "code-cushman-002": 2048,
            "code-cushman-001": 2048,
        }

        # handling finetuned models
        if "ft-" in modelname:
            modelname = modelname.split(":")[0]

        context_size = model_token_mapping.get(modelname, None)

        if context_size is None:
            raise ValueError(
                f"Unknown model: {modelname}. Please provide a valid OpenAI model name."
                "Known models are: " + ", ".join(model_token_mapping.keys())
            )

        return context_size

def __new__(cls, **data: Any) -> Union[OpenAIChat, BaseOpenAI]:  # type: ignore
        """Initialize the OpenAI object."""
        model_name = data.get("model_name", "")
        if (
            model_name.startswith("gpt-3.5-turbo") or model_name.startswith("gpt-4")
        ) and "-instruct" not in model_name:
            warnings.warn(
                "You are trying to use a chat model. This way of initializing it is "
                "no longer supported. Instead, please use: "
                "`from langchain_community.chat_models import ChatOpenAI`"
            )
            return OpenAIChat(**data)
        return super().__new__(cls)

Part4

使用方法

pip install openai

export OPENAI_API_KEY="..."

from langchain_openai import OpenAI

llm = OpenAI(openai_api_key="...")

llm.invoke(
    "What are some theories about the relationship between unemployment and inflation?"
)

'\n\n1. The Phillips Curve Theory: This suggests that there is an inverse relationship between unemployment and inflation, meaning that when unemployment is low, inflation will be higher, and when unemployment is high, inflation will be lower.\n\n2. The Monetarist Theory: This theory suggests that the relationship between unemployment and inflation is weak, and that changes in the money supply are more important in determining inflation.\n\n3. The Resource Utilization Theory: This suggests that when unemployment is low, firms are able to raise wages and prices in order to take advantage of the increased demand for their products and services. This leads to higher inflation.'

from langchain_openai import ChatOpenAI

chat = ChatOpenAI(openai_api_key="...")

from langchain_core.messages import HumanMessage, SystemMessage

messages = [
    SystemMessage(content="You're a helpful assistant"),
    HumanMessage(content="What is the purpose of model regularization?"),
]

chat.invoke(messages)

AIMessage(content="The purpose of model regularization is to prevent overfitting in machine learning models. Overfitting occurs when a model becomes too complex and starts to fit the noise in the training data, leading to poor generalization on unseen data. Regularization techniques introduce additional constraints or penalties to the model's objective function, discouraging it from becoming overly complex and promoting simpler and more generalizable models. Regularization helps to strike a balance between fitting the training data well and avoiding overfitting, leading to better performance on new, unseen data.")

Part5

其他