💻 Computer Engineering Q&A Chatbot

A deterministic, embedding-based question–answer retrieval system for core Computer Engineering topics, built using sentence embeddings and cosine similarity.

Author: MA Dehghan
Project Type: Linear Algebra Assignment
Last Updated: January 01, 2026

---

📑 Table of Contents

• Project Overview
• Demo
• Unique Features
• Tech Stack
• How It Works
• Mathematical Foundation
• Installation and Setup
• How to Run
• Adding or Customizing Questions
• Limitations
• Contributing
• Acknowledgments

---

🔍 Project Overview

This project implements a semantic question–answer retrieval system using vector embeddings and cosine similarity. Rather than generating answers, the chatbot retrieves the most semantically similar predefined question from a curated question bank.

The goal of the project is to demonstrate practical applications of linear algebra concepts — vector spaces, normalization, dot products, and similarity metrics — in natural language processing.

---

🎨 Demo

Example Interaction:

• User Query: "What is Moore's Law?"
• Chatbot Response: "Moore's Law states that the number of transistors on a microchip doubles approximately every two years, leading to exponential growth in computing power..."

If no strong match (similarity ≤ 0.3), it responds:

"I don't know!"

---

✨ Unique Features

---

🛠️ Tech Stack

• Language: Python 3.8+ 🐍
• Web Framework: Streamlit 🌟
• Embedding Model: Sentence Transformers - all-MiniLM-L6-v2 🤗 (Model Hub)
• Core Libraries: PyTorch ⚡, NumPy 🔢, Transformers 📦, JSON 📄
• Additional Utilities: Torch Functional (normalization, pooling); Streamlit Session State & Caching

---

📚 How It Works

1. Initialization: Load question bank from questionBank.json. Cache tokenizer/model.
2. User Input Handling: Accept query via Streamlit chat; append to session history.
3. Embedding Computation: Tokenize, run through model, mean pooling, L2 normalization.
4. Similarity Matching: Dot product with bank entries; select highest score. If ≤ 0.3 → "I don't know!".
5. Response Delivery: Split answer into words; stream to user via generator.

---

Install dependencies:

pip install streamlit torch transformers numpy

---

🧮 Mathematical Foundation

Embeddings

• Questions are embedded into a 384-dimensional vector space (ℝ³⁸⁴) using the all-MiniLM-L6-v2 model.
• Token-level embeddings are aggregated using mean pooling.
• Resulting vectors are L2-normalized to unit length.

Similarity Metric

Cosine similarity is used to measure semantic similarity between the user query and stored questions:

$\cos(\theta) = \frac{u \cdot q}{\lVert u \rVert \lVert q \rVert}$

Where:

• q = query embedding
• u = stored question embedding

Because all vectors are L2-normalized, cosine similarity reduces to a dot product between vectors, enabling efficient similarity computation.

---

🛠️ Installation and Setup

Prerequisites

Python 3.8+

Git

Install Dependencies pip install streamlit torch numpy transformers sentence-transformers

---

▶️ How to Run

git clone https://github.com/Mhalexmd/ChatBot
cd ChatBot
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
streamlit run app.py

Access at http://localhost:8501.

---

➕ Adding or Customizing Questions

questionBank.json format:

{
  "questions": [
    {
      "question": "What is a CPU?",
      "answer": "The central processing unit, executing instructions and managing data flow.",
      "embedding": [0.12, -0.34, ...]
    }
  ]
}

Generate embeddings with Python:

from sentence_transformers import SentenceTransformer
import json

model = SentenceTransformer('all-MiniLM-L6-v2')
questions = [{"question": "Your question", "answer": "Your answer"}]

for q in questions:
    q['embedding'] = model.encode(q['question']).tolist()

with open('questionBank.json', 'w') as f:
    json.dump({"questions": questions}, f)

---

⚠️ Limitations

Only answers in question bank

No multi-turn context

Slow for >1k questions without indexing

English-only for now

---

🤝 Contributing

Fork → Branch → Pull Request. Focus: New CE questions, optimizations, bug fixes.

---

🙏 Acknowledgments

Hugging Face, Streamlit, PyTorch

Linear algebra applications in NLP

Assignment inspiration