enhanced-ai-sdk

A lightweight utility that makes it easy to generate structured objects from LLMs using the Vercel AI SDK models and Zod schemas — with robust parsing powered by BAML under the hood.

Why: Many LLMs don’t natively support structured output, which makes Vercel AI SDK’s generateObject unusable with them. This library bridges that gap so those models can still produce typed objects directly.

• Familiar API: generateObjectCompat({ model, schema, ... }), similar in spirit to Vercel AI SDK's generateObject
• Strong typing via Zod
• Works with any Vercel AI SDK–compatible model (e.g. @ai-sdk/openai)
• Resilient parsing: converts your Zod schema to a temporary BAML schema and uses the generated parser to parse model output

Installation

This package has peer dependencies on the Vercel AI SDK and Zod.

# with pnpm
pnpm add enhanced-ai-sdk ai @ai-sdk/openai zod

# with npm
npm install enhanced-ai-sdk ai @ai-sdk/openai zod

# with yarn
yarn add enhanced-ai-sdk ai @ai-sdk/openai zod

Quick start

import { z } from 'zod';
import { createOpenAI } from '@ai-sdk/openai';
import { generateObjectCompat } from 'enhanced-ai-sdk';

const openai = createOpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  // Optional: point to compatible providers (e.g. DeepSeek via OpenAI-compatible gateway)
  baseURL: process.env.OPENAI_BASE_URL,
});

const schema = z.object({
  name: z.string(),
  color: z.string(),
});

const res = await generateObjectCompat({
  model: openai.chat('gpt-4o-mini'), // or any compatible model id
  schema,
  prompt: 'Generate a fruit with some description.',
});

console.log(res.object); // -> { name: string, color: string }
console.log(res.text); // raw model text

You can also pass messages (ChatML-style) instead of prompt.

API

export interface GenerateObjectCompatOptions<S extends z.ZodType = z.ZodType> {
  model: any; // a Vercel AI SDK model (e.g. from @ai-sdk/openai)
  schema: S; // Zod schema describing the expected object
  prompt?: string; // optional single prompt
  messages?: any[]; // optional chat messages
  // passthrough generation options
  maxTokens?: number;
  temperature?: number;
  topP?: number;
}

export interface GenerateObjectCompatResult<T> {
  object: T; // parsed object (z.infer<typeof schema>)
  text: string; // raw model output text
  usage?: any; // provider usage (if available)
  finishReason?: any; // provider finish reason (if available)
}

How it works

1. Your Zod schema is converted into a BAML schema on the fly.
2. A minimal JSON-only instruction is injected into your prompt/messages to encourage valid JSON output.
3. The LLM is called via Vercel AI SDK's generateText.
4. The raw text is parsed by the generated BAML client into the typed object.

The generated BAML client is cached in a temporary directory (e.g. ~/<tmp>/baml-runtime-cache).

Requirements

• Node.js 18+
• API credentials for your chosen model provider (e.g. OPENAI_API_KEY) and optional OPENAI_BASE_URL if using an OpenAI-compatible gateway.

Scripts

bun run build   # build the library with rslib
bun run dev     # build in watch mode
bun run test    # run tests with rstest (skips real model test if OPENAI_API_KEY is missing)
bun run format  # format with Biome
bun run check   # lint/check with Biome

Testing

There is an integration-style test that will make a real model call if OPENAI_API_KEY is set (and optionally OPENAI_BASE_URL). If the key is missing, the test is skipped.

OPENAI_API_KEY=sk-... bun run test

Caveats & notes

• While the library strongly nudges the model to output JSON and uses a robust parser, LLMs can still return malformed outputs in edge cases. Errors will be thrown if parsing fails.
• messages vs prompt: if you provide both, messages are used and the JSON-only system hint is merged into the first system message (or prepended if none exists).
• This package is not an official Vercel AI SDK package; it simply aims to provide a familiar, convenient developer experience for structured outputs.

License

MIT License © 2025 James Tsang. See LICENSE for full text.