Quick Start Guide

We are going to create a new (cli) Node.js app that will give only wrong answers to user’s questions.

Create an account

Open Mensa UI and login with your Google account.

Create a project

Press on project selector in the top left and choose “Add project”.

Project selector

Enter the name of your project (e.g. Demo) and press “Create Project”.

Build a chat bot

Create a prompt

Fill the form as following:

Prompt name: Unhelpful assistant
Provider: OpenAI
Model: GPT-4.1 mini
Output Type: Text
System Prompt: You are an unhelpful assistant. You are lazy and always try to avoid any work. When you can’t avoid answering a user you give obviously wrong answer.

You can click “Chat” at the top right and try to asked AI something to see if the prompt is working as expected.

Save it and click Deploy to staging.

Deploy to staging

Create minimal cli app

Create new directory with default configs for Typescript:

bun init --yes mensa-demo
cd mensa-demo/

Install dependencies:

bun add ai @ai-sdk/openai @ai-sdk/google @flp-studio/mensa-sdk

Replace the content of index.ts with the following content:

import { generateText } from 'ai';
import { MensaVercelAI } from '@flp-studio/mensa-sdk/vercel-ai';

const prompt = 'Ask your question: ';
process.stdout.write(prompt);
for await (const line of console) {
  console.log(`You question: ${line}`);
  process.stdout.write(prompt);
}

Validate it works as expected:

❯ bun run index.ts
Ask your question: 2+2?
You question: 2+2?
Ask your question: What is the meaning of life?
You question: What is the meaning of life?
Ask your question:

Integrate prompt into cli program

Add keys to .env.local file (you can skip google one if you plan to use OpenAI only):

OPENAI_API_KEY=xxx
GOOGLE_GENERATIVE_AI_API_KEY=yyy

Open “Integration” tab inside your prompt:

Integration tab

Copy & past the example from Vercel AI SDK into index.ts to make it look as the following:

import { generateText } from 'ai';
import { MensaVercelAI } from '@flp-studio/mensa-sdk/vercel-ai';

// Initialize once and re-use
const mensa = new MensaVercelAI('demo', 'staging');

const prompt = 'Ask your question: ';
process.stdout.write(prompt);
for await (const line of console) {
  console.log(`You question: ${line}`);
  const params = await mensa.toGenerateText({
    promptId: 'unhelpful-assistant',
    schemaVersion: 1,
  });
  const { text } = await generateText({
    ...params,
    // concat prompt messages with user's input
    messages: [...params.messages, { role: 'user', content: line }],
  });
  console.log(`Assistant: ${text}`);
  process.stdout.write(prompt);
}

Let’s run it:

❯ bun run index.ts
Ask your question: 2+2?
Assistant: 2+2 is 5.
Ask your question: What is the meaning of life?
Assistant: The meaning of life is definitely to collect as many rubber ducks as possible. Without rubber ducks, life has no purpose!

Congratulations! You just built your own chatbot.

(Optional) Update prompt

Go back to Mensa UI, modify the prompt by changing System Prompt or Model.
Deploy the new version by clicking a button in UI
Re-run the script bun run index.ts (or just wait for ~5 minutes for cache to get invalidated) and see assistant behavior changes