Simple Intent-Based AI Chatbot

A lightweight, neural network-powered chatbot that recognizes user intents using a bag-of-words approach + a small feed-forward neural network built with PyTorch.

Currently recognizes the following intents:

• greeting
• goodbye
• programming (explains what coding is)
• resource (recommends learning resources)
• stocks (shows a random selection from your "portfolio")

Features

• Intent classification using a 3-layer feed-forward neural network
• Bag-of-words (binary) input representation
• Tokenization + lemmatization with NLTK
• PyTorch model training & inference
• Supports function calling (e.g. stocks intent → calls get_stocks())
• Model saving & loading
• Simple command-line chat interface

Demo

Enter your message: hi
Hello!

Enter your message: how can i learn coding?
Check out the NeuralNine YouTube channel and The Python Bible series (7 in 1).

Enter your message: what are my stocks?
Here are your stocks!
['NVDA', 'META', 'MSFT']

Enter your message: bye
Goodbye!

Project Structure

chatbot/
├── main.py              # Main application + model + training loop
├── intents.json         # Training patterns + responses
├── chatbot_model.pth    # Trained model weights (after training)
├── dimensions.json      # Input/output size info for loading
└── README.md

Requirements

Python 3.8+
torch
numpy
nltk

Install dependencies:

uv add install torch numpy nltk

Then download required NLTK data (only needed once):

import nltk
nltk.download('punkt')
nltk.download('wordnet')
nltk.download('omw-1.4')   # optional but recommended

How to Use

Option 1: Use the pre-trained model

uv run main.py

Option 2: Retrain the model from scratch

Uncomment these lines in main.py

assistant = ChatbotAssistant('intents.json', function_mappings = {'stocks': get_stocks})
assistant.parse_intents()
assistant.prepare_data()
assistant.train_model(batch_size=8, learning_rate=0.001, epochs=100)
assistant.save_model('chatbot_model.pth', 'dimensions.json')

Then run:

uv run main.py

The model will be retrained and saved.

How to Add New Intents

1. Open intents.json
2. Add a new intent object like this:

   {
   "tag": "thanks",
   "patterns": ["Thanks", "Thank you", "Thx", "Appreciate it"],
   "responses": ["You're welcome!", "Anytime!", "Glad I could help"]
   }

3. (Optional) Add a function mapping in main.py:

   function_mappings = {
       'stocks': get_stocks,
       'thanks': lambda: print("Extra action: sending virtual high-five!")
   }

4. Retrain the model (see Option 2 above)

Current Limitations & Possible Improvements

• Very small dataset → performance is limited on unseen phrases
• Binary bag-of-words → no word order or frequency information
• No handling of out-of-scope questions (falls back to predicted intent)
• No context/memory between messages

Future ideas

• Use TF-IDF instead of binary BoW
• Add word embeddings (e.g. pretrained GloVe / fastText)
• Implement a larger model (LSTM / Transformer)
• Add entity extraction (names, dates, products…)
• Connect to real APIs (weather, news, real stock data…)
• Create a web / Discord / Telegram interface

License

• MIT License – feel free to use, modify, and share.