Swapping Into LangChain

The payoff from langchain Lesson 11

The langchain course's Lesson 11 (init_chat_model) made a promise: LangChain's whole point is that swapping AI providers means changing a line or two, not rewriting your program. This lesson cashes that promise in for real, against a genuinely different kind of provider, a local model instead of a second cloud one.

ChatOllama, from the langchain-ollama package, is a LangChain chat model class, the same family as ChatGoogleGenerativeAI used everywhere else in this repo. Same .invoke(), same .bind_tools(), same AIMessage objects back. The only thing that changes is which class you instantiate and what's underneath it.

The code, piece by piece

model = ChatOllama(model="llama3.2", temperature=0)

Compare this to ChatGoogleGenerativeAI(model="gemini-3.5-flash-lite") from langchain Lesson 1: same shape, no API key needed here for the reason covered back in Lesson 3 of this course, there's no remote provider to authenticate against.

response = model.invoke("In one sentence, what is Python used for?")
response.text

Identical to every other LangChain lesson in this repo. response is a real AIMessage, .text pulls out the written reply, exactly the same as calling Gemini.

@tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    ...

model_with_tools = model.bind_tools([get_weather])
tool_response = model_with_tools.invoke("What is the weather in Paris?")

This is the langchain course's @tool and .bind_tools() pattern (Lessons 13-14 there), completely unchanged. LangChain translates the decorated function into whatever tool-calling format Ollama actually expects (the same shape as Lesson 11's raw TOOLS dict in this course), so you never write that translation yourself.

Why reach for this instead of the raw `ollama` package

Lessons 3-12 used the ollama package directly, and that's a perfectly good choice for a program that only ever needs to talk to Ollama. ChatOllama earns its place when you want the option to swap providers later without a rewrite (development against a free local model, production against a more capable cloud one, say), or when your agent is already built with LangChain's abstractions (chains, LangGraph nodes, MCP tool adapters) and Ollama needs to slot into that same shape.