Customer Service Chatbot

AI-powered customer service chatbot using DeepSeek (LLM) + SQLite (context database). Built with Node.js, Express, and vanilla frontend styled with Tailwind CSS.

Tech Stack

Layer	Tech
Backend	Node.js + Express
AI	DeepSeek API (deepseek-chat)
Database	SQLite (better-sqlite3)
Frontend	Vanilla HTML + CSS + JS
Styling	Tailwind CSS (CDN)
Markdown	marked.js (CDN)

Project Structure

.
├── .env                  # Environment variables
├── .gitignore
├── package.json
├── server.js             # Express entry point, routes, static serve
├── cache.js              # In-memory TTL response cache
├── db.js                 # SQLite connection, LIKE search, full-table fetch
├── context.js            # Tokenizer, intent detection, context builder
├── ai.js                 # DeepSeek API client + caching layers
└── public/
    ├── index.html        # Chat UI
    ├── css/style.css     # Custom styles (scrollbar, markdown, typing anim)
    └── js/app.js         # Frontend chat client

Setup

npm install

Create a .env file:

DEEPSEEK_API_KEY=sk-your_key_here
DEEPSEEK_BASE_URL=https://api.deepseek.com
DEEPSEEK_MODEL=deepseek-chat

DB_PATH=./data.db
PORT=3000
CONTACT_EMAIL=cs@example.com

DS_TEMPERATURE=0.7
DS_TOP_P=0.9
DS_TOP_K=40
DS_MAX_TOKENS=2048
DS_PRESENCE_PENALTY=0
DS_FREQUENCY_PENALTY=0

Place your SQLite database file at the path specified in DB_PATH, then:

npm start

Open http://localhost:3000.

Database Schema

The chatbot expects a SQLite database with these tables:

abouts

Column	Type	Description
section_title	TEXT	Section heading (e.g. "Brand", "Vision")
content	TEXT	Body content

faqs

Column	Type	Description
question	TEXT	Frequently asked question
answer	TEXT	Corresponding answer

posts

Column	Type	Description
title	TEXT	Blog/article title
content	TEXT	Blog/article body

products

Column	Type	Description
name	TEXT	Product name
short_description	TEXT	Brief product summary
long_description	TEXT	Detailed product description
tags	TEXT	Comma-separated tags
link	TEXT	Product slug or URL
price	TEXT	Price as string

stores

Column	Type	Description
name	TEXT	Store name
address	TEXT	Store address

API

POST /api/chat

Send a user message and get an AI reply grounded in database context.

Request:

{
  "message": "Do you have aloe vera gel?"
}

Response:

{
  "reply": "Yes! We have Aloe Vera Gel — a soothing gel made from 92% aloe vera, priced at $12.99.",
  "hasData": true
}

Field	Type	Description
reply	string	AI-generated response
hasData	boolean	Whether relevant database context was found

Status codes:

• 200 — Success
• 400 — Missing message field
• 500 — Internal server error (AI API failure, etc.)

Frontend

The public/ directory is served as static files. The chat UI sends POST /api/chat requests via fetch() and renders responses with markdown.

How Context Works

1. Tokenize — User message is split into keywords (stop words removed)
2. LIKE search — Each keyword is matched against all 5 tables via SQL LIKE
3. Intent fallback — If a table has zero results but the message hints at it (e.g. "store", "toko", "price", "harga"), all rows from that table are fetched
4. Context assembly — Results are formatted into a text block prefixed by section headers
5. AI prompt — The context is injected into DeepSeek's system prompt; the AI must answer using only that data

Caching Strategy

Two layers:

Layer	What it caches	TTL	Avoids
App cache (cache.js)	Exact Q&A response (by MD5 hash of message + context)	1 hour	Duplicate API calls entirely
DeepSeek auto cache	Prompt prefix (system + context) on server side	Automatic (≥1024 tokens)	Resending cached prompt tokens

Console logs show cache efficiency: [App Cache] HIT, [DS Auto Cache] hit=X miss=Y.

Environment Variables

Variable	Default	Description
DEEPSEEK_API_KEY	—	DeepSeek API key (required)
DEEPSEEK_BASE_URL	https://api.deepseek.com	API base URL
DEEPSEEK_MODEL	deepseek-chat	Model name
DB_PATH	./data.db	Path to SQLite database
PORT	3000	HTTP server port
DS_TEMPERATURE	0.7	Response randomness
DS_TOP_P	0.9	Nucleus sampling
DS_TOP_K	40	Top-K sampling
DS_MAX_TOKENS	2048	Max response length
DS_PRESENCE_PENALTY	0	Penalize repeated topics
DS_FREQUENCY_PENALTY	0	Penalize repeated phrases

Language Support

English and Indonesian. Stop words and intent keywords are defined for both languages:

• Intent tables map common query words to database tables:
  • store / toko → stores
  • product / produk → products
  • about / tentang → abouts
  • faq / retur → faqs
  • blog / artikel → posts