Skip to main content

Chains

Chains are a key concept in LangChain that allow combining multiple components together to build complex applications. At a high level, a chain takes some input, passes it through a series of components like LLMs and parsers, and returns the final output.

Why Chains are Needed​

Chains are needed for several reasons:

  • Modularity - They allow breaking down complex logic into reusable components
  • Simplified Debugging - Each component can be tested in isolation
  • Flexibility - Components can be swapped in and out easily
  • Prompt Engineering - Chains make it easy to shape the context and prompt seen by the LLM

Without chains, applications would need large monolithic functions that are hard to maintain.

Getting Started with Chains​

The simplest chain is the LLMChain. It combines:

  • A language model (LLM)
  • A prompt template
  • Optionally, an output parser

Here is a simple example:

from langchain import PromptTemplate, OpenAI, LLMChain

prompt = PromptTemplate(
  template="What is a good name for a company that makes {product}?"
)

llm = OpenAI() 

chain = LLMChain(llm=llm, prompt=prompt)

print(chain.run("colorful socks"))

This shows the power of chains - it abstracts away the prompt formatting, querying the LLM, and parsing the output into one simple call.

ChatModels vs LLMs​

A key difference is that ChatModels take a list of ChatMessage objects as input rather than just a string. For example:

from langchain import ChatOpenAI, HumanMessage

chat_model = ChatOpenAI()

messages = [
  HumanMessage(content="What is a good name for a company that makes colorful socks?")  
]

response = chat_model.predict_messages(messages)
print(response.content)

Understanding the difference between LLMs and ChatModels is important for constructing effective prompts.

Summary​

In summary, chains simplify combining components like LLMs, prompt templates, and parsers to build robust applications. They provide modularity, improved debugging, flexibility, and prompt engineering benefits.