Streaming Objects

It's nice being able to stream text from an LLM, but sometimes you want that output to be more structured. For instance, you may want it to provide multiple queries that you search Google with.

In other words, sometimes we want text back from the LLM, but sometimes we want objects. We want to define the shapes of those objects, pass that shape to the LLM, and then get that shape back.

The Problem

Our starting code calls streamText to generate a story about an imaginary planet:

const stream = streamText({
  model,
  prompt: "Give me the first paragraph of a story about an imaginary planet.",
})

We then write that to stdout and await the final text:

for await (const chunk of stream.textStream) {
  process.stdout.write(chunk)
}

const finalText = await stream.text

The Challenge: Stream a Structured Object

Our first to-do is to call streamObject from the AI SDK, passing in:

1. The model (Gemini 2.0 Flash)
2. A prompt asking for facts about the imaginary planet
3. A Zod schema defining the structure we want back

Here's where we need to add our code:

// TODO: Replace this with a call to streamObject, passing:
// - The model, same as above
// - The prompt, asking for facts about the imaginary planet,
//   passing in the finalText as the story
// - The schema, which should be an object with a facts property
//   that is an array of strings
const factsResult = TODO

The schema we need to define should be an object with a facts property that's an array of strings. LLMs are very good at working with Zod schemas, so autocomplete should help you define this properly.

Handling the Stream

After we get our structured data stream, we'll log each chunk of the partial object stream:

for await (const chunk of factsResult.partialObjectStream) {
  console.log(chunk)
}

This will show the object as it's being constructed by the LLM in real-time.

When you run this exercise, you should see:

1. The text of the story streaming in first
2. A series of object logs showing the current shape of the facts object as it's being generated

The final result should be an array of facts about the imaginary planet described in the paragraph that was generated.

Steps To Complete

• Call the streamObject function and store the result in factsResult

  • Import streamObject from 'ai' if not already imported

• Pass the model parameter to streamObject (same as used with streamText)

• Create a prompt that asks for facts about the imaginary planet

  • Include the finalText in your prompt so the LLM knows what story to reference

• Define a Zod schema for an object with a facts property that is an array of strings

  • Import z from 'zod' if not already imported
  • Use the z.array and z.string functions to define the schema
  • Use the z.describe function to describe the schema, if you like

• Run the exercise with pnpm run dev to see both the streaming text and the structured object being built