Swapping Into Pydantic AI

No dedicated Ollama model class, and that's fine

Unlike ChatOllama in Lesson 13, pydantic_ai has no OllamaModel class of its own. Instead, Ollama exposes an OpenAI-compatible endpoint at http://localhost:11434/v1, meaning any client built for OpenAI's API shape can talk to it directly. pydantic_ai already ships an OpenAIChatModel; OllamaProvider just points that same class at your local server instead of api.openai.com.

model = OpenAIChatModel(
    model_name="llama3.2",
    provider=OllamaProvider(base_url="http://localhost:11434/v1"),
)

This is a genuinely useful pattern to recognize beyond Ollama: whenever a model provider doesn't have first-class support in a framework you're using, check whether it exposes an OpenAI-compatible endpoint. Many do (this is exactly the interoperability pattern that makes cheap swapping between providers possible in the first place).

Three agents, three pydantic_ai patterns, one local model

plain_agent = Agent(model, system_prompt="You are a concise assistant.")
plain_result = plain_agent.run_sync(...)

An ordinary Agent, unstructured text output, identical usage to the pydantic_ai course's earliest lessons.

structured_agent = Agent(model, output_type=NativeOutput(Movie))

output_type=Movie alone (pydantic_ai's usual default) asks the model to call a hidden "return this result" tool, which works well against strong cloud models but can be less reliable on smaller local ones. NativeOutput(Movie) routes around that: it uses the provider's own JSON-schema-constrained generation instead, the exact format mechanism from Lesson 10, just reached through pydantic_ai's API instead of the raw ollama package. Worth knowing this option exists specifically because local models sometimes need it.

tool_agent = Agent(model, system_prompt="You are a weather assistant.")

@tool_agent.tool_plain
def get_weather(city: str) -> str:
    ...

@agent.tool_plain, unchanged from the pydantic_ai course. The same manual ask/call/respond loop from Lesson 11 of this course, just hidden behind pydantic_ai's decorator instead of written out by hand.