import dspy
from typing import Literal

Using `Literal` locks the output to valid categories only — the AI can't hallucinate labels. `ChainOfThought` adds reasoning which improves accuracy over bare `Predict`.

class TagContent(dspy.Signature):
    """Assign all applicable tags to the content."""
    message: str = dspy.InputField(desc="The content to tag")
    tags: list[Literal[tuple(CATEGORIES)]] = dspy.OutputField(desc="All applicable tags")

tagger = dspy.ChainOfThought(TagContent)

from dspy.evaluate import Evaluate

def sorting_metric(example, prediction, trace=None):
    return prediction.category == example.category

evaluator = Evaluate(
    devset=devset,
    metric=sorting_metric,
    num_threads=4,
    display_progress=True,
    display_table=5,
)
score = evaluator(sorter)

optimizer = dspy.BootstrapFewShot(
    metric=sorting_metric,
    max_bootstrapped_demos=4,
)
optimized_sorter = optimizer.compile(sorter, trainset=trainset)

optimizer = dspy.MIPROv2(
    metric=sorting_metric,
    auto="medium",
)
optimized_sorter = optimizer.compile(sorter, trainset=trainset)

result = optimized_sorter(message="I was charged twice on my credit card last month")
print(f"Category: {result.category}")
print(f"Reasoning: {result.reasoning}")