Skip to main content

LLMChain

LLMChain allows chaining together prompts and models to build simple yet powerful AI applications. This guide provides an overview of LLMChain and examples for getting started.

Introduction​

LLMChain is one of the core building blocks in LangChain. It chains together a prompt template and a language model to format prompts, query models, and return responses.

Some key features:

  • Supports both LLMs and chat models
  • Easy formatting of prompts using input variables
  • Flexible input methods including dicts, lists, and keywords
  • Additional options like apply for batch predictions

Constructing an LLMChain​

To use LLMChain, first create a PromptTemplate and a model.

from langchain.llms import OpenAI 
from langchain.prompts import PromptTemplate

llm = OpenAI(temperature=0.9)
prompt = PromptTemplate(
  input_variables=["product"],
  template="What is a good name for a company that makes {product}?"  
)

Then construct the chain by passing the prompt and model.

from langchain.chains import LLMChain

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

Passing Inputs​

There are a few ways to pass inputs to an LLMChain:

Dict

print(chain.run({"product": "colorful socks"}))

Keywords 

print(chain.predict(product="colorful socks"))

List

inputs = [{"product": "socks"}, {"product": "shoes"}]
print(chain.apply(inputs))

Additional Prediction Methods​

LLMChain provides a few additional ways to run predictions:

  • apply - Runs on a batch of inputs and returns a list of outputs
  • generate - Returns a LLMResult with token usage and other metadata
  • predict - Specifies inputs as keywords rather than a dict

For example:

inputs = [{"product": "socks"}, {"product": "computer"}]

outputs = chain.apply(inputs) # Returns list of outputs
result = chain.generate(inputs) # Returns LLMResult

chain.predict(product="shoes") # Keyword arguments

Conclusion​

In summary, LLMChain is a simple yet powerful building block that chains together prompts and models. It provides flexibility through multiple input methods, batch predictions, access to metadata, and easy composability into more complex chains.