从开发应用到训练模型：能力边界、资源成本和选择顺序

OPTIONS = {
    "prompt": {"data": 1, "compute": 1, "ops": 1, "freshness": 2},
    "rag": {"data": 2, "compute": 2, "ops": 3, "freshness": 5},
    "sft": {"data": 4, "compute": 4, "ops": 4, "freshness": 1},
    "alignment": {"data": 5, "compute": 5, "ops": 5, "freshness": 1},
}


def rank(requirement: dict) -> list[tuple[str, int]]:
    scores = {}
    for name, cost in OPTIONS.items():
        penalty = (
            cost["data"] * requirement["data_scarcity"]
            + cost["compute"] * requirement["compute_limit"]
            + cost["ops"] * requirement["ops_limit"]
            + abs(cost["freshness"] - requirement["freshness_need"])
        )
        scores[name] = penalty
    return sorted(scores.items(), key=lambda item: item[1])


print(rank({
    "data_scarcity": 3,
    "compute_limit": 3,
    "ops_limit": 2,
    "freshness_need": 5,
}))