AI FAQ assistant with guardrails

A shop assistant that answers questions with an LLM, stays on topic, declines everything else and never sends a fixed answer to the model.

Level: Beginner · Tags: AI, WhatsApp, Telegram, Guardrails

Features shown: AI autoresponder · business-scoped persona · text and function triggers · slash commands · /reset of the AI history · cost and latency settings · terminal simulator

Prerequisites

• Node.js 22.12 or newer.
• An Anthropic API key in the ANTHROPIC_API_KEY environment variable (or another model provider, see below). Never write the key in the code.

npm install @botmesh/core @botmesh/ai @anthropic-ai/sdk @botmesh/baileys

The bot

// bot.js
import { Bot, defineBot, isMainModule, slashCommandsPlugin } from '@botmesh/core'
import { BaileysProvider } from '@botmesh/baileys'
import { aiPlugin } from '@botmesh/ai'

// The persona is the main guardrail: one business, what to answer, what to decline.
const PERSONA = [
  'You are the assistant of Northwind Bikes, a bicycle shop with a repair workshop.',
  'Answer only about our bikes, accessories, repairs, opening hours (Mon-Sat 9:00-19:00) and orders.',
  'If the question is about anything else, say in one sentence that you can only help with Northwind Bikes',
  'and suggest typing "agent" to talk to a person.',
  'Never make up prices, stock or policies: if you do not know, say so. Answer in at most three short sentences.',
].join(' ')

const createBot = defineBot((provider) => {
  const bot = new Bot({ provider })

  bot.plug(slashCommandsPlugin, {
    greeting: '🚲 Hi! I am the Northwind Bikes assistant. Ask me about bikes, repairs or opening hours.',
  })

  // Fixed answers never reach the model: instant, free and always right.
  bot.when(['hours', 'opening hours'], (ctx) => ctx.reply('🕘 We are open Mon-Sat, 9:00-19:00.'))
  bot.when(['agent', 'human'], (ctx) => ctx.reply('👤 Call us at [phone] or write to [e-mail].'))

  // A deterministic guardrail: very long messages are not sent to the model.
  bot.when(
    (ctx) => ctx.body.length > 500,
    (ctx) => ctx.reply('✂️ That is a long message. Please ask one short question at a time.'),
  )

  // Everything else goes to the model. Plugged last, so its /reset also clears the AI history.
  bot.plug(aiPlugin({
    provider: 'anthropic',
    apiKey: process.env.ANTHROPIC_API_KEY,
    model: 'claude-haiku-5-5',     // a small model is enough for FAQs
    persona: PERSONA,
    maxTokens: 300,                // short answers: cheaper and faster
    context: { maxMessages: 10 },  // history sent with each question
  }))

  return bot
})

export default createBot

// `node bot.js` connects to WhatsApp; `npx botmesh dev bot.js` skips this block.
if (isMainModule(import.meta)) {
  const bot = await createBot(new BaileysProvider({ session: 'faq-bot' }))
  bot.on('ready', () => console.log('✅ Connected'))
  await bot.start()
}

How it works

• Order of handlers. Commands, bot.when() triggers and flows always run before the AI. aiPlugin registers a catch-all autoresponder that only sees the text messages nothing else handled, so “hours” and “agent” are answered by your code, not by the model.
• The persona is the guardrail. It names one business, lists what the bot may talk about and says how to decline. Since January 2026 WhatsApp only allows AI in service of a business, never as a general-purpose assistant (see WhatsApp AI policy).
• A deterministic guardrail. A function trigger answers messages longer than 500 characters itself, so a pasted essay never becomes a paid model call. Add any rule you can check in code the same way.
• Cost and latency. model, maxTokens and context.maxMessages keep every answer small. timeoutMs (60 s by default) bounds each call; a failed call answers the built-in “⚠️ I can’t answer right now…” text instead of crashing.
• /reset. aiPlugin adds its own /reset, which clears the chat’s AI history. It is plugged after slashCommandsPlugin, so its /reset wins.

Try it

npx botmesh dev bot.js --env-file .env   # .env has the line ANTHROPIC_API_KEY=sk-ant-...

This is a run of the code above in the terminal simulator. The AI answers come from a scripted model in the test that runs this page (packages/ai/tests/docs-examples-ai.test.ts); a real model words them differently.

user › /start
bot › 🚲 Hi! I am the Northwind Bikes assistant. Ask me about bikes, repairs or opening hours.
user › hours
bot › 🕘 We are open Mon-Sat, 9:00-19:00.
user › do you repair electric bikes?
bot › Yes, our workshop services electric bikes too: motor, battery and brakes. Bring it in Mon-Sat, 9:00-19:00.
user › write me a poem about the sea
bot › Sorry, I can only help with Northwind Bikes. Type "agent" to talk to a person.
user › agent
bot › 👤 Call us at [phone] or write to [e-mail].

Other model providers

Switch to OpenAI, Gemini or a local Ollama model

Install the SDK of the provider and change the aiPlugin options; the rest of the bot stays the same.

OpenAI (npm install openai, key in OPENAI_API_KEY):

aiPlugin({ provider: 'openai', apiKey: process.env.OPENAI_API_KEY, model: 'gpt-5-mini', persona: PERSONA })

Gemini (npm install @google/genai, key in GEMINI_API_KEY):

aiPlugin({ provider: 'gemini', apiKey: process.env.GEMINI_API_KEY, model: 'gemini-3.8-flash', persona: PERSONA })

Ollama on your machine (no package, no key; run ollama pull llama3.1 first):

aiPlugin({ provider: 'ollama', model: 'llama3.1', persona: PERSONA })

The full list (and the OpenAI-compatible presets) is in Choose a provider.

Run it on WhatsApp or Telegram

node bot.js prints a QR code: scan it with WhatsApp (Linked devices) and write to that number from another phone. For Telegram, install @botmesh/telegram and start the factory with new TelegramProvider({ token: process.env.TELEGRAM_TOKEN ?? '' }) instead (see Or connect Telegram).

Next steps

• AI with tools: let the model book appointments, with a human confirmation
• AI plugin guide: every aiPlugin option
• Examples overview