Return typed triage with PydanticAI

Define the decisions as Pydantic fields. The native integration turns each field into a question and returns an instance of your output model.

pip install 'gemmadecision[pydantic-ai]'

The local model uses the shared CPU ONNX runtime by default.

from typing import Literal

from pydantic import BaseModel, Field
from pydantic_ai import Agent

from gemmadecision import DecisionEngine, GemmaDecisionModel


class Triage(BaseModel):
    """Decide how a support ticket should enter the queue."""

    team: Literal["billing", "account", "technical", "review"] = Field(
        description=(
            "Which team should handle the ticket? Billing handles payments; "
            "account handles access; technical handles software problems; "
            "review handles unclear requests."
        )
    )
    urgent: bool = Field(
        description="Does the ticket give a deadline today or say work is currently blocked?"
    )


def build_triage_agent(engine: DecisionEngine):
    return Agent(GemmaDecisionModel.local(engine=engine), output_type=Triage)


def main() -> None:
    engine = DecisionEngine.from_pretrained()
    agent = build_triage_agent(engine)
    result = agent.run_sync(
        "The application crashes when I upload a file, and I cannot finish today's report."
    )
    print(result.output.model_dump_json(indent=2))
    print(result.response.provider_details)


if __name__ == "__main__":
    main()

Reuse the engine and agent. The result has .team and .urgent fields with the declared types. The questions are evaluated together, with one decision per field; this is not a guarantee that all fields satisfy application-specific relationships. Validate those relationships separately.

For the shortest setup, use GemmaDecisionModel.local() without passing an engine. It loads the shared default model on first use. In an async application, use await agent.run(ticket) instead of run_sync().

Finite Literal, Enum and boolean fields fit this API. A free-form str field such as summary is unsupported because this model does not generate text. The PydanticAI guide covers the available schemas.