NUST Admissions Guide - Offline Chatbot

An offline, RAG-powered admissions chatbot for NUST (National University of Sciences and Technology) that runs entirely on student hardware — no internet, no GPU, no cloud APIs.

Built for the NUST Islamabad Local Chatbot Competition 2026.

Demo Video

Watch the full demo on YouTube

Screenshots

Main Interface with System Monitor

Answering Eligibility Criteria

NET Exam Information

Honest Uncertainty — Says "I Don't Know" Instead of Hallucinating

Architecture

User Question
      │
      ▼
┌─────────────┐     ┌──────────────────┐
│  Embedding  │────▶│  FAISS Vector DB │
│  (MiniLM)   │     │  (1070 chunks)   │
└─────────────┘     └──────┬───────────┘
                           │ Top-5 chunks
                           ▼
                    ┌──────────────────┐
                    │   Qwen2.5-3B     │
                    │  (Q4_K_M GGUF)   │
                    │   CPU Inference   │
                    └──────┬───────────┘
                           │
                           ▼
                    ┌──────────────────┐
                    │  Gradio Chat UI  │
                    └──────────────────┘

Data Sources

All data is sourced from official NUST channels:

• 74 FAQ Q&A pairs scraped from nust.edu.pk/faqs/
• 12 official NUST PDFs (prospectus, fee policy, eligibility criteria, NET subjects, hostel rates, etc.)
• Official admissions pages (merit criteria, admission procedure, fee structure)

Design Tradeoffs

Decision	Choice	Why
LLM	Qwen2.5-3B Q4_K_M	Best quality-to-size ratio for CPU; 2.0GB fits in 8GB RAM
Embeddings	all-MiniLM-L6-v2	Only 80MB, fast on CPU, excellent semantic search
Vector Store	FAISS	Zero-dependency, fast similarity search, no server needed
Quantization	Q4_K_M	Sweet spot between quality and speed
Context	2048 tokens	Enough for RAG context + answer; keeps inference fast
Retrieval	Top-5 chunks	Balances context richness vs. inference speed
UI	Gradio	Clean, responsive, zero-config, works offline

Hardware Requirements

• RAM: 4-6 GB (model ~2.0GB + embeddings ~80MB + OS overhead)
• CPU: Any modern CPU (optimized for i5 13th gen)
• GPU: Not required
• Storage: ~3 GB for model + data
• Internet: Only needed for initial setup (downloading model)

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Download the LLM Model

python download_model.py

This downloads Qwen2.5-3B (Q4_K_M, ~2.0GB) from HuggingFace.

3. Build the Knowledge Base

python ingest.py

This processes the admissions data and builds the FAISS vector index.

4. Launch the Chatbot

python app.py

Opens at http://127.0.0.1:7860

Key Features

• 100% Offline: After initial setup, no internet needed
• Official Data Only: All answers sourced from nust.edu.pk and official PDFs
• Transparent: Shows source documents and response time for every answer
• Honest: Explicitly says "I don't have this information" instead of hallucinating
• Fast: Optimized CPU inference with quantized model and batch processing
• Extensible: Add new data by dropping files in data/raw/ and re-running python ingest.py

Project Structure

ChatBot/
├── app.py              # Gradio UI with system monitor
├── rag.py              # RAG pipeline (retrieval + generation)
├── ingest.py           # Data ingestion and vector store builder
├── config.py           # All configuration in one place
├── download_model.py   # Model downloader
├── benchmark.py        # Performance benchmarking
├── requirements.txt    # Python dependencies
├── screenshots/        # Demo screenshots
├── data/
│   ├── raw/            # Official NUST text data (scraped)
│   ├── pdfs/           # Official NUST PDF documents
│   └── vector_store/   # FAISS index (auto-generated)
└── models/
    └── model.gguf      # LLM model (auto-downloaded)

---

Built by Muhammad Mujeeb Awan (MSEE-2K24, SEECS) for the NUST Islamabad Local Chatbot Competition 2026 Runs entirely offline on student hardware (8GB RAM, Core i5, no GPU)