Streaming with streamText

streamText returns immediately and delivers the answer as it is generated. textStream yields pieces of text; the result's text and usage resolve when the stream ends.

Examplestream.ts

import { streamText } from "ai";
import { shopModel } from "./shop-model.ts";

const result = streamText({ model: shopModel, prompt: "My parcel never arrived" });

for await (const piece of result.textStream) {
  process.stdout.write(`[${piece}]`);
}
console.log();
console.log(await result.text);
console.log((await result.usage).totalTokens);

Example

npx tsx stream.ts

streamText is not awaited: it starts the request and returns a result whose streams fill as the model writes. Each [piece] above is one chunk, a word, as the stand-in sends them. result.text and result.usage are promises that resolve once the stream is done.

Every event

Exampleevents.ts

import { streamText } from "ai";
import { shopModel } from "./shop-model.ts";

const result = streamText({
  model: shopModel,
  prompt: "I was charged twice",
  onEnd: ({ finishReason }) => console.log("\nfinished:", finishReason),
});

for await (const part of result.fullStream) {
  process.stdout.write(`${part.type} `);
}

Example

npx tsx events.ts

fullStream yields every part, not only text: the start and end of the call and of each step, text boundaries, and, with tools, tool calls and results. onEnd runs after the last part. A UI uses these to show "looking up your order" while a tool runs.

Consume the stream

The request only makes progress while something reads the stream. A route that returns result.toUIMessageStreamResponse() lets the browser read it; a script has to loop over it or await result.text.

Try it yourself

Print part.text for text-delta parts only.
Add onChunk and count chunks.
Remove the loop and only await result.text.

PreviousPrompts, system instructions and messages

Little by little, you're building something great.