CallEats: AI-Powered Restaurant Voice Assistant

Multi-tenant voice assistant system for restaurants using Vapi.ai, semantic search, and real-time call management.

Overview

AI-powered voice assistant for restaurants that automates phone inquiries using Vapi.ai and semantic search. Handles menu questions, hours, delivery zones, and availability through natural conversation.

Key Capabilities

Capability	Description
Multi-Tenant Voice Assistant	Shared Vapi.ai assistant routes calls to correct restaurants using phone number mapping
Semantic Knowledge Base	Vector similarity search using OpenAI embeddings (text-embedding-3-small, 1536 dimensions) and pgvector
Real-Time Call Management	Automatic call tracking with fallback mechanisms for unreliable webhooks
Intelligent Caching	Redis-based distributed caching (60s TTL) with in-memory fallback
Production Architecture	FastAPI backend, React frontend, comprehensive middleware stack

System Architecture

graph TB
    subgraph "Customer"
        A[Phone Call]
    end

    subgraph "Voice AI"
        B[Vapi.ai Assistant]
        C[GPT-4o-mini]
    end

    subgraph "Backend API"
        D[FastAPI Server]
        E[Phone Mapping]
        F[Vector Search]
        G[Call History]
        H[Cache Layer]
    end

    subgraph "Data Layer"
        I[(Supabase<br/>PostgreSQL)]
        J[pgvector<br/>Embeddings]
        K[(Redis Cache)]
    end

    subgraph "External Services"
        L[OpenAI<br/>Embeddings API]
    end

    subgraph "Frontend"
        M[React Dashboard]
    end

    A --> B
    B --> C
    B --> D
    D --> E
    D --> F
    D --> G
    D --> H
    E --> I
    F --> J
    F --> L
    G --> I
    H --> K
    H --> I
    M --> D

    style A fill:#3b82f6
    style B fill:#8b5cf6
    style D fill:#10b981
    style F fill:#f59e0b
    style I fill:#06b6d4
    style M fill:#ef4444

Loading

Technical Approach

Multi-Tenant: All data scoped by restaurant_id with RLS policies and isolated vector search.

Voice Integration: Vapi.ai function tools (get_menu_info, get_modifiers_info, get_hours_info, get_zones_info) route to restaurant-specific data.

Vector Search: OpenAI embeddings → pgvector similarity search → cached results (60s TTL).

Call Tracking: Webhook events + scheduled API fetch fallback for reliability.

Technology Stack

Component	Technology	Purpose
Backend	FastAPI, Python 3.12	REST API with async support
Voice AI	Vapi.ai, GPT-4o-mini	Voice assistant and conversation
Embeddings	OpenAI text-embedding-3-small	Semantic search vectorization
Vector DB	Supabase pgvector	Vector similarity search
Database	Supabase PostgreSQL	Multi-tenant data storage
Cache	Redis / cachetools	Distributed caching with fallback
Frontend	React, TypeScript, Vite	Restaurant management dashboard
State Management	React Query (TanStack Query)	Server state and caching
Styling	Tailwind CSS	Responsive UI components
Error Tracking	Sentry	Production error monitoring
Authentication	JWT, Supabase Auth	User authentication and authorization

Key Features

• Dashboard: Real-time call stats, recent history, quick actions
• Menu Builder: Category/item management, modifiers, image uploads
• Call History: Complete transcripts, metadata, filtering
• Delivery Zones: Geographic zones with Leaflet maps, fee configuration
• Operating Hours: Day-of-week scheduling, open/close times

Infrastructure

Service	Platform	Purpose
Backend	Railway	FastAPI service
Frontend	Vercel	React SPA
Database	Supabase	PostgreSQL + pgvector
Cache	Railway Redis	Distributed cache (in-memory fallback)
Voice	Vapi.ai	Voice assistant platform
Monitoring	Sentry	Error tracking

Getting Started

Prerequisites

• Python 3.12+
• Node.js 18+
• Supabase account
• Vapi.ai account
• OpenAI API key
• Redis (optional, for distributed caching)

Backend Setup

cd backend
pip install -r requirements.txt
cp env.example .env
# Configure environment variables
uvicorn restaurant_voice_assistant.main:app --reload

Frontend Setup

cd frontend
npm install
npm run dev

Database Migration

Run migrations in order:

1. 001_initial_schema.sql - Core schema with pgvector
2. 002_phone_mappings.sql - Phone number routing
3. 003_users_table.sql - User authentication
4. 004_delivery_zones_geometry.sql - Geographic zones
5. 005_menu_items_image_url.sql - Image support

Vapi Configuration

Use scripts/setup_vapi.py to configure Vapi.ai assistant with function tools and webhook endpoints.

Use Cases

Menu inquiries, pricing questions, hours information, delivery coverage, modifier questions.

Key Differentiators

1. Multi-tenant architecture with shared assistant instance
2. Semantic search via vector similarity
3. Production-ready with error handling, rate limiting, monitoring
4. Webhook reliability with API fetch fallback
5. Real-time embedding updates on menu changes
6. Scalable Redis caching with graceful fallback

Performance

• Search: 200-500ms (cached), 1-2s (cache miss)
• Embeddings: 100-300ms per query
• Vector search: 50-200ms (HNSW index)
• Cache hit rate: 60-80%
• Call data fetch: 2-5s

Security

JWT auth, webhook secret validation, RLS policies, rate limiting, request validation, security headers.

Documentation

• Architecture Documentation
• API Reference

---

Status: Production-ready portfolio project demonstrating multi-tenant voice AI architecture