Task Results

Validate model output with Python types and application rules

result_type specifies the Python value a task should produce. It defaults to str and can be a scalar, collection, dataclass, Pydantic model, or set of classification labels.

Structured values

import marvin
from pydantic import BaseModel, Field

class Review(BaseModel):
    summary: str
    rating: int = Field(ge=1, le=5)

review = marvin.run(
    "Summarize and rate the supplied feedback",
    result_type=Review,
    context={"feedback": "Fast delivery, but the packaging was damaged."},
)

The model receives the result schema. Invalid results produce validation feedback and can be retried. A valid type or rating range does not establish that the content is correct.

Classification

Pass labels, a Literal, or an enum to constrain the result:

from typing import Literal
import marvin

status = marvin.run(
    "Classify the supplied deployment status",
    result_type=Literal["healthy", "degraded", "down"],
    context={"status": "All health checks passed."},
)

For a classification helper and multi-label examples, see classify.

Custom validation

A result_validator must return the value to store or raise an exception to reject it. It replaces the task’s ordinary result-validation path, so it must perform any validation or conversion your application needs.

import marvin

def positive_count(value: int) -> int:
    if value <= 0:
        raise ValueError("Count must be positive")
    return value

task = marvin.Task(
    "How many items are in a pair?",
    result_type=int,
    result_validator=positive_count,
)
count = task.run()

For work that only performs actions, use result_type=None. See running tasks for task states and failure handling.