Vercel AI SDK

Use Schema Stream with the Vercel AI SDK when you need finer updates than partialOutputStream provides, schema-derived placeholders before fields arrive, or completion events for individual JSON values. Feed streamText().textStream to the parser and keep result.output as the authoritative structured result.

import { type LanguageModel, Output, streamText } from "ai"
import { SchemaStream } from "schema-stream"
import { z } from "zod"

const analysisSchema = z.object({
  summary: z.string(),
  details: z.object({ score: z.number() }),
  tags: z.array(z.string())
})

async function streamAnalysis({
  input,
  model
}: {
  input: string
  model: LanguageModel
}): Promise<z.output<typeof analysisSchema>> {
  const result = streamText({
    model,
    output: Output.object({ schema: analysisSchema }),
    prompt: input
  })
  const parser = new SchemaStream(analysisSchema)

  for await (const snapshot of parser.iterate(result.textStream)) {
    renderProgress(snapshot)
  }

  return await result.output
}

The LanguageModel input keeps the parsing code independent of how the application configures an AI SDK provider or gateway. Schema Stream integrates with the normalized AI SDK text stream; it does not configure provider credentials or replace the SDK's structured-output validation.

Choose the native partial stream when it is enough

Prefer AI SDK's partialOutputStream when its object cadence already meets the interface's needs. Use Schema Stream on textStream when one or more of these matter:

• incomplete strings should visibly grow;
• nested placeholders should exist before the provider reaches them;
• onValueComplete should trigger a conditional branch as soon as one value settles;
• snapshot cadence should use Schema Stream's chunk, value, bytes, or final policy.

Regardless of the progress path, await result.output before treating the generation as valid.

Verify the integration

examples/sdk-mocks.ts and tests/sdk-runtime.test.ts exercise streamText(), Output.object(), and textStream with provider-shaped deterministic events:

bun run example:sdk
bun test tests/sdk-runtime.test.ts

The test suite covers Unicode, escaped content, optional and nullable fields, records, and deeply nested arrays. The packed-consumer gate also compiles the integration against the generated package:

bun run test:packed

An opt-in Vercel AI Gateway lane exists for an application-selected provider/model id:

SCHEMA_STREAM_LIVE_E2E=1 \
SCHEMA_STREAM_LIVE_PROVIDER=gateway \
SCHEMA_STREAM_GATEWAY_MODEL=<provider/model> \
bun run test:live

It requires AI_GATEWAY_API_KEY through the server environment. See Provider portability before generalizing the deterministic AI SDK coverage to a specific Anthropic or Gemini model.