LangChain

A Python/JS framework for composing LLM calls, prompts, tools, and memory into end-to-end applications.

At a glance

Field	Value
Category	LLM application framework
Difficulty	Intermediate
When to use	Prototyping pipelines with many integrations
When not to use	Production systems where you want minimal deps
Alternatives	LlamaIndex, DSPy, Haystack, raw SDKs

What it is

LangChain provides primitives — Runnable, ChatPromptTemplate, RetrievalQA, output parsers — and a long list of pre-built integrations for every LLM provider, vector DB, and loader under the sun. The LCEL pipe syntax (prompt | model | parser) composes these into chains.

When we reach for it at Ephizen

Early prototypes that need to talk to OpenAI, Anthropic, and a vector DB in the same afternoon.
Reference implementations of standard RAG patterns you don’t want to rewrite.
Interop with the broader LangChain ecosystem (LangSmith for tracing, LangGraph for agents).

Getting started

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate

prompt = ChatPromptTemplate.from_messages([
    ("system", "You answer questions from context."),
    ("human", "Context:\n{context}\n\nQ: {question}"),
])
chain = prompt | ChatOpenAI(model="gpt-4o-mini")
chain.invoke({"context": docs, "question": "What's our refund policy?"})

Gotchas

LangChain has broken its APIs multiple times. Pin versions aggressively.
The framework leaks abstractions — debugging a chain often means reading LangChain’s source.
For serious agent work, LangGraph (same team, state-machine based) is a saner foundation than the original agent classes.
Many teams eventually strip LangChain out in favor of direct SDK calls once the shape of the app stabilizes.