Vercel AI SDK¶

Director-AI ships a TypeScript bridge for the Vercel AI SDK middleware API. It uses wrapLanguageModel() and reviews model output through the Director-AI REST /v1/review endpoint.

cd packages/vercel-ai
npm install
npm run build

import { generateText, wrapLanguageModel } from "ai";
import { createDirectorAiMiddleware } from "@director-ai/vercel-ai";

// `yourModel` is any AI SDK language model instance.
const guardedModel = wrapLanguageModel({
  model: yourModel,
  middleware: createDirectorAiMiddleware({
    endpoint: "https://director.example.com",
    apiKey: process.env.DIRECTOR_API_KEY,
  }),
});

const result = await generateText({
  model: guardedModel,
  prompt: "What is the refund policy?",
});

Streaming Safety¶

The middleware buffers streamText() text deltas, reviews the final text, and only releases chunks after Director-AI approves the output. This preserves the AI SDK stream protocol while avoiding a retroactive block after unsafe tokens have already reached the client.

Reject Handling¶

By default, rejected output raises DirectorAiGuardrailError without embedding the raw model response in the error message.

createDirectorAiMiddleware({
  endpoint: "https://director.example.com",
  onReject: "throw",
});

Use mask mode when the application should return a fixed suppression message:

createDirectorAiMiddleware({
  endpoint: "https://director.example.com",
  onReject: "mask",
  rejectionMessage: "This answer was blocked by the safety layer.",
});

Contract¶

Option	Purpose
`endpoint`	Director-AI server root, for example `http://localhost:8080`
`apiKey`	Optional `X-API-Key` value
`reviewPath`	Override path, default `/v1/review`
`onReject`	`"throw"` or `"mask"`
`rejectionMessage`	Mask-mode replacement text

The bridge targets AI SDK v6 LanguageModelV3Middleware.