ZapCaf — AI Café Chatbot

An AI-powered chatbot for a fictional Indo-Western café, built from scratch using PyTorch and NLTK. Understands natural language queries about menu, delivery, payments, store hours, and more — deployed as a web app with a Flask backend and a clean browser-based UI.

Live Demo → realtime-chatbot-gz8y.onrender.com

---

What It Does

A user types a message like "What time do you open?" or "Do you accept UPI?" — the system classifies the intent behind the message using a trained neural network and returns the appropriate response. If it's not confident enough, it gracefully says it doesn't understand.

---

System Design & Data Flow

Training Pipeline (Offline — run once)

intents.json
     │
     │  json.load()
     ▼
[tag, patterns, responses]        ← 13 intents, 70+ pattern sentences
     │
     │  nltk.word_tokenize()
     ▼
Token lists per pattern
     │
     │  PorterStemmer.stem() + filter punctuation + sorted(set())
     ▼
Vocabulary: 54 unique stemmed words   ← the feature space
     │
     │  bag_of_words()  →  binary vector per pattern
     ▼
x_train: (N, 54) float32             ← input features
y_train: (N,)    int                 ← class labels (0–12)
     │
     │  ChatDataset + DataLoader (batch=8, shuffle=True)
     ▼
NeuralNet  →  54 → 8 → 8 → 13
     │
     │  CrossEntropyLoss + Adam (lr=0.001) × 1000 epochs
     ▼
data.pth   ← saved model weights + vocabulary + tags

---

Inference Pipeline (Runtime — per message)

User types message
     │
     │  tokenize()          →  word list
     │  stem()              →  root forms
     │  bag_of_words()      →  54-dim binary vector
     ▼
torch.tensor  shape (1, 54)
     │
     │  NeuralNet.forward()
     ▼
Raw logits  shape (1, 13)
     │
     │  torch.max()         →  predicted class index
     │  tags[index]         →  predicted intent tag
     │  torch.softmax()     →  confidence probability
     ▼
confidence > 0.75 ?
     ├── YES → random.choice(intent["responses"])
     └── NO  → "I do not understand..."
     │
     ▼
Response string

---

Web Architecture

Browser
  │
  │  GET /          →  Flask serves index.html
  │  POST /predict  →  JSON {message} → JSON {answer}
  ▼
Flask (app_server.py)
  │
  │  get_response(msg)
  ▼
chat.py  →  nltk_utils.py  →  model.py
  │
  └──  data.pth  (loaded once at startup)

---

Tech Stack

Layer	Technology
Neural Network	PyTorch
NLP	NLTK (tokenization), PorterStemmer (stemming)
Feature Engineering	Bag-of-Words
Web Backend	Flask
Frontend	HTML, CSS, JavaScript
Deployment	Render.com

---

Project Structure

RealTime_ChatBot/
│
├── intents.json          # Training data — intents, patterns, responses
├── nltk_utils.py         # tokenize(), stem(), bag_of_words()
├── model.py              # NeuralNet definition (PyTorch)
├── train.py              # Training script → outputs data.pth
├── data.pth              # Saved model + vocabulary
├── chat.py               # Inference engine — get_response()
│
├── app_server.py         # Flask web server
├── app.py                # Original Tkinter desktop app
│
├── static/
│   ├── style.css         # All CSS
│   ├── script.js         # Chat UI logic
│   ├── zappy.png         # Hero image
│   └── zap_Robo.png      # Chat mascot
│
├── templates/
│   └── index.html        # Web frontend
│
├── render.yaml           # Render deployment config
├── requirements.txt      # Python dependencies
└── QnA.md                # Interview Q&A reference

---

Neural Network Architecture

Input Layer     →  54 neurons  (bag-of-words vocabulary size)
Hidden Layer 1  →   8 neurons  + ReLU
Hidden Layer 2  →   8 neurons  + ReLU
Output Layer    →  13 neurons  (one per intent class)

Total trainable parameters: 54×8 + 8 + 8×8 + 8 + 8×13 + 13 = 645
Loss function: CrossEntropyLoss
Optimizer: Adam (lr=0.001)
Epochs: 1000

---

Intents Covered

Intent	Example Query
greeting	"Hi", "Hello", "Good morning"
goodbye	"Bye", "See you later"
thanks	"Thanks", "Much appreciated"
items	"What do you sell?"
payments	"Do you accept UPI?"
delivery	"How long does delivery take?"
funny	"Tell me a joke!"
store_hours	"What time do you open?"
menu	"Show me the menu"
reservations	"Can I book a table?"
complaints	"My order was wrong"
location	"Where are you located?"
offers	"Any discounts today?"

---

Run Locally

# Clone
git clone https://github.com/Adyasha56/RealTime_ChatBot.git
cd RealTime_ChatBot

# Install dependencies
pip install -r requirements.txt

# Train the model (only needed if intents.json changes)
python train.py

# Start the web server
python app_server.py

Open http://127.0.0.1:10000 in your browser.

---

Retrain the Model

To add new intents, edit intents.json following the existing structure, then run:

python train.py

This regenerates data.pth. No other files need to change.

---

License

MIT License — free to use, modify, and distribute.