AI Guest Communication for Hotels: Voice + Chat + Messenger

A sole build: an n8n-orchestrated stack for four UK hospitality venues across embedded web chat, Facebook Messenger, and AI phone reception (Vapi + Twilio). Same retrieval-backed answers and lead capture everywhere, with venue-specific knowledge and guardrails so guests get accurate answers rather than a generic hotel bot.

The problem

Guests message and call outside staff hours, and high-value event enquiries sit next to "where do I park?" and get lost. Teams need 24/7 coverage, consistent answers per venue, and qualified leads with context, without doubling headcount.

What I shipped

• Three channels: website widget (HTTP), Meta Messenger (signed webhooks + send API), and voice (inbound Twilio to a Vapi conversational layer).
• RAG over venue data: embeddings + PostgreSQL / PGVector per venue (menus, packages, policies, FAQs) so replies stay grounded in that site's facts.
• Lead capture: for events, private hire, and groups, structured fields, urgency hints, source channel, and logging to Google Sheets for the sales team.
• Session memory: Postgres-backed conversation state so follow-ups ("any vegetarian options?") stay in context.
• Business rules: table bookings redirect to the venue's real booking flow, and capture triggers are tuned for high-intent threads, not every chat.
• Hosting & ops: self-hosted n8n on DigitalOcean with HTTPS for webhooks; I owned deploy and ongoing tweaks.

Architecture (at a glance)

 Web widget       Messenger          Phone (Twilio)
     | HTTP         | webhook            | SIP
     v              v                    v
              n8n workflows (orchestration)
     |               |                     |
     v               v                     v
OpenAI chat +   PGVector retrieval     Vapi voice layer
 embeddings      (per-venue KB)         + human transfer
     |
     v
Postgres session memory  ·  Google Sheets (lead CRM)

Traffic from web, Messenger, and Twilio fans into n8n workflows that call OpenAI (chat + embeddings) and PGVector retrieval, then write optional Sheets rows for leads. Voice adds Vapi on the telephony path with human transfer when needed.

Stack

n8n (self-hosted), OpenAI GPT-4.1-mini + embeddings, PostgreSQL + PGVector, Vapi + Twilio, Meta Graph API (Messenger), Google Sheets, DigitalOcean.

Quality & result

Before go-live I ran an automated eval suite (factual checks per venue, regression cases from real feedback, and multi-turn lead flows), scoring 36/36 on the final Bellini run with CSV exports for audit. In production, venues get round-the-clock guest handling and structured leads instead of missed off-hours DMs and voicemails.