AIordering (AI Food Ordering Agent)

Project origin: This repository originates from the main branch of JunCaiAgent/AIordering. This project mainly delivers feature extensions and architecture improvements on top of the original (frontend/backend separation, authentication system, admin console, recommendation enhancements, production deployment support, etc.).

refactor branch — Frontend/Backend Separation + Production Deployment Enhancements

A restaurant food-ordering agent built on LangChain 1.0 + FastAPI + LLMs. Users interact through natural language to query dishes, get smart recommendations by party size / taste / diner profile, and manage dishes & users via an admin console. This branch upgrades the first version (main branch) with four major areas: frontend/backend separation, an authentication system, a full admin backend, and production deployment support.

Differences from the First Version (main branch)

Aspect	First version (main)	This version (refactor)
Project layout	Flat files, single FastAPI app serving everything	Frontend/backend separation: frontend/ (H5 pages) + backend/ (API service)
Recommendation	Text-to-SQL + basic recommendations	RAG vector retrieval (ChromaDB, 40 dish profiles) + rules engine + diversified recommendation algorithm (category balancing / dynamic weights / soft-constraint fallback — never returns an empty result)
Database	restaurant schema	catering_agent schema (renamed & migrated, 27 dishes, new pic image field)
Authentication	No login system	User registration/login (SMS/password) + image captcha + token auth; all business APIs require login
Admin console	None	Full admin backend (/admin): overview stats, dish pagination & inline editing, delist/relist, user management (create users / set passwords), recommendation weight & rule config, KB rebuild, data-statistics charts
Dish delisting	None	is_active flag: delisted dishes are excluded from rebuilt KB and recommendations
Deployment	Local run only	Gunicorn production config, BT-Panel deployment, nginx reverse proxy, static-asset performance (93% image size reduction)

Features

• Natural-language ordering chat: dish lookup, recommendation, pairing advice (/api/ai/chat)
• Smart recommendation: quota by party size, taste fallback chain, diner/scene filters, health tags, hard constraints on allergens & dietary taboos, explicit counts ("recommend one signature dish"), context awareness ("girlfriend is on her period" → warm sweet drinks)
• Image captcha: Pillow-rendered, Arial Bold with stroke, confusable characters removed from the charset, case-insensitive
• User auth: SMS/password login, 30-day tokens, set/reset password
• Admin console: dashboard, dish pagination & inline edit, user accounts, weight/rule config, KB rebuild (with amber warning banner), statistics (sales ranking / donut chart / trends)
• Demo ordering: mock order endpoint (login required, persists to DB and feeds back sales counts)

Tech Stack

• Backend: Python 3.11 / FastAPI / LangChain 1.0 / Uvicorn / Gunicorn
• Vector store: ChromaDB (1024-dim, embedded by qwen3.7-text-embedding)
• Databases: MySQL 8 (local) / 5.7 (server), Redis (session & rate limiting)
• LLMs: Alibaba Cloud Bailian qwen3.7-flash (chat), qwen3.7-text-embedding (embedding — model names must be lowercase)
• Frontend: vanilla H5 single-page, Font Awesome icons

Project Layout

├── frontend/            # H5 frontend (served by nginx / FastAPI)
│   ├── index.html       # User ordering chat
│   ├── admin.html       # Admin console
│   └── img/             # Dish images (27, compressed)
└── backend/             # Backend service
    ├── main/            # api_server / agent / tools / auth / admin / db ...
    ├── kb_data/         # ChromaDB knowledge base (rebuilt by build_kb.py, git-ignored)
    ├── config.py        # Configuration (reads .env)
    ├── build_kb.py      # Rebuild knowledge base
    ├── weight_calc.py   # Recommendation weight calculation
    ├── gunicorn_conf.py # Production multi-worker config
    ├── requirements.txt
    └── .env.example     # Env template (real .env holds secrets, git-ignored)

Quick Start (local development)

# 1. Install dependencies (Python 3.11)
cd backend && pip install -r requirements.txt

# 2. Configure environment
cp .env.example .env    # fill in DB / Bailian LLM / SMS credentials

# 3. Initialize database (schema + 27 dishes)
python main/init_db.py

# 4. Compute weights & rebuild KB
python weight_calc.py && python build_kb.py

# 5. Run
python main/api_server.py
# Open http://127.0.0.1:3000/ (user) and http://127.0.0.1:3000/admin (admin)

Default Accounts

When using the backend/catering_agent.sql snapshot from this repo, the following accounts are built in (passwords shown in plaintext):

Role	Account	Password	Entry
Admin	admin	test123456	Admin console /admin
Demo user	19900001111	test123456	User login (password)
Demo user	19900002222	test123456	User login (password)

Phone numbers are anonymized, non-real numbers for demo only; in production create users via the admin console and change the admin password.

Production Deployment

• Single-host: gunicorn -c gunicorn_conf.py main.api_server:app
• Separated + nginx reverse proxy: serve static frontend via nginx, proxy location /api/ to the backend on port 3000 (see deploy.sh / project docs)
• BT-Panel: Python Project Manager + nginx site reverse proxy (on CentOS 7: install deps with --only-binary=:all:; use pysqlite3-binary for ChromaDB on systems with sqlite < 3.35)

Author

李俊颖 and project contributors (Citadel Yang)