Skip to main content

What is LangChain?

PyPI Python License

A raw call to an LLM provider's API gives you one thing: text in, text out. Nearly every real application needs more than that — a system prompt assembled from templates, a way to give the model tools it can call, documents to search before answering, a running conversation, and a way to see what actually happened when something goes wrong. LangChain is the composition layer that supplies those pieces so you don't rebuild them per project.

Key idea

LangChain is a composition layer over model providers, not a model itself. It standardizes how prompts, models, tools, retrieval, and output parsing plug into each other so the same code can swap a provider, add a tool, or add retrieval without a rewrite.

What it gives you​

  • A uniform model interface — init_chat_model and the chat model classes expose the same invoke / stream / batch surface regardless of which provider is behind them.
  • Composable pipelines (LCEL) — prompt, model, and parser chain together with |, and the resulting pipeline is itself invokable, streamable, and batchable.
  • Retrieval building blocks — loaders, splitters, embeddings, and vector store integrations for retrieval-augmented generation.
  • Tool calling and agents — a standard way to describe tools to a model and route the model's tool-call requests back to your code.
  • Observability — traces of every step in a chain or agent run, via LangSmith.

What you still write​

LangChain does not decide your product's prompts, your retrieval strategy, your data model, or your error-handling policy. It gives you typed building blocks; the judgment calls — which chunks to retrieve, when an agent should stop, what counts as a safe tool call — stay yours.

The stack​

Your application code calls into LangChain's abstractions; LangChain calls into each provider's SDK; the SDK talks to the provider's API. Retrieval branches off to a vector store instead of (or alongside) the model call.

See also​