Documentation menuVercel AI SDK

Vercel AI SDK

Mark every answer and every generated image automatically with AI SDK middleware.

import { openai } from "@ai-sdk/openai";
import { generateImage, streamText, wrapImageModel, wrapLanguageModel } from "ai";
import { witnessImageMiddleware, witnessMiddleware } from "@sweberdev/witness/ai-sdk";

const model = wrapLanguageModel({
  model: openai("gpt-5"),
  middleware: witnessMiddleware({ provider: "Beispiel AG" }),
});

const result = streamText({ model, prompt });

witnessMiddleware:

appends the text watermark to every text part, in generateText and at the end of every text part in streamText;
adds providerMetadata.witness with aiGenerated, kind, sourceType, generator and createdAt;
calls onMarked for each marked part, so you can log it.

Option	Default
`specificationVersion`	`v4`	`v4` for `ai` 7, `v3` for `ai` 6, `v2` for `ai` 5
`generator`	`provider/modelId`	String or function of the model
`kind`	`generated`
`provider`	none	Your company or product
`watermark`	`true`	`false` keeps only the metadata
`id`	none	Function returning a reference written into the watermark
`onMarked`	none	Called with `{ type, marking, partId }`

witnessImageMiddleware writes image XMP into every image from generateImage, whether the provider returns base64 or bytes. Images with a C2PA manifest are left as they are, and unsupported formats produce a warning on the result instead of an error.

const imageModel = wrapImageModel({
  model: openai.image("gpt-image-2"),
  middleware: witnessImageMiddleware({ provider: "Beispiel AG" }),
});
const { image } = await generateImage({ model: imageModel, prompt: "Mountain lake at dawn" });

The module has no dependency on ai; it uses the middleware shapes structurally.

Show the notice in the chat UI as well: the middleware marks content, it does not inform the person in front of the screen. See Chatbot notice.