Generate Enums With Vercel's AI SDK

Another classic use case for LLMs is classification.

Let's say we want to pass a user comment to the LLM to classify as positive, negative, or neutral.

We want to get back from the LLM essentially one string, and that string is going to be either positive, negative, or neutral.

In traditional software parlance, this is called an enum — a set of enumerated values.

And we can use the AI SDK to generate that enum.

Scrollycoding

We're using the generateObject function again, but this time we're passing an output of enum.

We also pass an enum property with an array of the possible strings we want it to return.

The result we get back from generateObject has an object property, and that object is our enum.

import { generateObject } from "ai";

export const classifySentiment = async (
  text: string,
) => {
  await generateObject({
    model,
    output: "enum",
    prompt: text,
    system:
      `Classify the sentiment of the text as either ` +
      `positive, negative, or neutral.`,
  });
};

Let's try out a few different statements to see if it works:

Scrollycoding

I'm not sure how I feel comes out as neutral.

This is terrible comes out as negative.

And I love this so much comes out as positive.

const result = await classifySentiment(
  `I'm not sure how I feel`,
);

console.log(result); // neutral

And just like that, we've got a sentiment analysis system. This is a really cool use case for enums, and it's great that the AI SDK makes it so simple.