Building a WhatsApp chatbot in Haitian Creole with an LLM

Use the official WhatsApp Business Platform (Cloud API). Unofficial libraries that drive WhatsApp Web get numbers banned.

User ──WhatsApp──► Meta ──webhook──► your server ──► LLM (+ your FAQ/docs)
                                          │
User ◄──WhatsApp── Meta ◄──send message───┘

// 1. Webhook verification (one-time GET from Meta)
app.get("/whatsapp/webhook", (req, res) => {
  if (req.query["hub.verify_token"] === process.env.WA_VERIFY_TOKEN) return res.send(req.query["hub.challenge"]);
  res.sendStatus(403);
});
 
// 2. Incoming messages
app.post("/whatsapp/webhook", async (req, res) => {
  res.sendStatus(200); // acknowledge fast; Meta retries slow webhooks
  const msg = req.body.entry?.[0]?.changes?.[0]?.value?.messages?.[0];
  if (!msg || msg.type !== "text") return;
 
  const history = await loadRecentTurns(msg.from, 10);   // short memory per phone number
  const answer = await askModel({
    system: "Ou se asistan sipò [Sèvis]. Reponn an kreyòl ayisyen, kout (maks 4 fraz). " +
            "Sèvi sèlman ak enfòmasyon ki nan DOKIMAN yo. Si ou pa konnen, di sa epi bay nimewo sipò a.",
    docs: await retrieveRelevantFaq(msg.text.body),        // RAG over your FAQ
    history,
    user: msg.text.body,
  });
  await sendWhatsAppText(msg.from, answer);
  await saveTurn(msg.from, msg.text.body, answer);
});

Practical points:

• Verify webhook signatures (the X-Hub-Signature-256 header) before processing.
• 24-hour window: you can reply freely within 24 hours of the user's last message. Starting a conversation later requires pre-approved templates.
• Keep answers short. Long paragraphs are painful on small screens and cost more.
• Offer a "pale ak yon moun" (talk to a human) escape, and hand off when the model isn't confident.
• Rate-limit per phone number, so one user or a spam loop can't burn your AI budget.