Generate Arrays With Vercel's AI SDK

So far we've looked at getting an LLM to return objects or enums (enumerated values) instead of just returning text.

But what if you want to return multiple objects - an array of objects? What then?

For instance, you might want your LLM to generate lots of fake data.

Zod Schema

Scrollycoding

Let's first create a Zod schema that we first encountered in our structured outputs example.

If you don't understand this I've got a free course on Zod on my sister site Total TypeScript.

We should also use descriptions on the schema fields to give the AI more context. Like we've seen before, we can use .describe here.

import { z } from "zod";

const schema = z.object({
  name: z.string(),
  age: z.number(),
  email: z.string().email(),
});

Passing The Schema To `generateObject`

Then we can pass this schema to the generateObject function - but we also pass in an output of array.

import { generateObject } from "ai";

export const createFakeUsers = async (
  input: string,
) => {
  const { object } = await generateObject({
    model,
    prompt: input,
    system: `You are generating fake user data.`,
    output: "array",
    schema,
  });

  return object;
};

Now the object that we get back from generateObject will be this array of users.

What's powerful about this is that we can seed various information about the users, so in this case, they're going to be from the UK like me.

Let's give this a go. We're going to generate some fake users:

const fakeUsers = await createFakeUsers(
  "Generate 5 fake users from the UK.",
);

console.dir(fakeUsers, { depth: null });

And just like that, we get our users.

This is particularly cool with streamObject too, where you can stream in the users as they're being created.

So that's how to generate an array of structured objects using the AI SDK.

Generate Arrays With Vercel's AI SDK

Zod Schema

Passing The Schema To generateObject

Passing The Schema To `generateObject`