Skip to main content

Language Models in LangChain

LangChain provides interfaces and integrations for two main types of language models: LLMs (Large Language Models) and Chat Models. Understanding the difference between these two types of models is key to using LangChain effectively.

LLMs​

LLMs (Large Language Models) are models that take a text string as input and return a text string as output. Examples include GPT-3, GPT-3.5, Codex.

To use an LLM in LangChain:

from langchain.llms import OpenAI

llm = OpenAI()

output = llm.predict("What is the capital of France?")

The predict method takes a string prompt and returns the model's generated string response.

Chat Models​

Chat models are tuned for having conversations. Their provider APIs take a list of chat messages as input rather than a single string. The messages are labeled with a speaker role, like "Human", "AI", or "System". The model returns a response chat message.

Examples of chat models include Claude, GPT-4, and Anthropic's conversational model.

To use a chat model in LangChain:

from langchain.chat_models import ChatOpenAI
from langchain.schema import HumanMessage

chat_model = ChatOpenAI() 

messages = [HumanMessage("What is the capital of France?")]
response = chat_model.predict_messages(messages)

The predict_messages method takes a list of ChatMessage objects and returns a ChatMessage response.

Chat Messages​

When working with chat models, inputs and outputs are ChatMessage objects. These contain:

  • content: The text content of the message
  • role: The speaker role, like "Human" or "AI"

LangChain provides helper classes for common roles:

  • HumanMessage: Message from a human
  • AIMessage: Message from an AI assistant
  • SystemMessage: Message from the system

Summary​

In summary, LLMs take string input and output while Chat Models take chat messages as input and output. LangChain provides interfaces for both, but understanding this key difference is important for using the library effectively. The helper classes like HumanMessage and AIMessage make working with Chat Models easy.