LangChain is a framework for building applications with large language models (LLMs). It started as a set of procedural chain classes but evolved into LCEL (Language Chain Expression Language), a declarative way to compose components. Runnables are the foundation: they're objects with a consistent interface (.invoke(), .batch(), .stream()) for chaining operations. This post walks through the architecture, showing how to think in terms of Runnables and when to use them.
Runnables: The Core Abstraction
A Runnable is any object with .invoke() (or .batch(), .stream()) methods. Runnables can be chained with the | operator.
from langchain.prompts import PromptTemplate
from langchain_openai import ChatOpenAI
# Create components (all Runnables)
template = "You are a helpful assistant. Answer this: {question}"
prompt = PromptTemplate(input_variables=["question"], template=template)
model = ChatOpenAI(model="gpt-4")
# Chain with | operator (LCEL)
chain = prompt | model
# Invoke
response = chain.invoke({"question": "What is 2+2?"})
print(response.content)
Output:
2 + 2 equals 4.
The | operator chains Runnables. The output of the left side becomes the input to the right side.
Old API vs. New API
Old API: chains were Python classes with custom logic.
# Old (not recommended)
from langchain.chains import LLMChain
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gpt-4")
prompt = PromptTemplate(input_variables=["question"], template="Q: {question}\nA:")
chain = LLMChain(prompt=prompt, llm=llm)
response = chain.run(question="What is 2+2?")
New API: Runnables are simpler and more composable.
# New (recommended)
prompt | ChatOpenAI(model="gpt-4")
The new approach is more concise and composable. Prefer it for new code.
Conclusion
LangChain's architecture centers on Runnables—composable objects with a consistent interface. LCEL makes building complex workflows simple and declarative. Understanding the Runnable protocol enables you to extend LangChain with custom components. Next: we'll explore Prompt Templates and Output Parsers—how to structure requests and parse responses.