🧠 RMIT Academic Policies Chatbot

AI-Assisted Supporter Chatbot for New RMIT Students

This chatbot helps new RMIT students understand academic policies such as academic integrity, assessment extensions, appeals, and course progress.
It provides accurate, citation-backed answers based on official RMIT policies using Retrieval-Augmented Generation (RAG) with AWS Bedrock (Claude Sonnet / Haiku).

---

✨ Features

• 🎨 Streamlit-based modern UI (mobile-friendly)
• ⚙️ RAG with FAISS + sentence-transformers
• 🧠 Context-aware conversation memory
• 💬 Automatic topic/entity clarification
• 📚 Auto-indexing of uploaded policy JSON files
• 🧩 Responsible-AI guardrails
• 📖 Clause-level citations for every factual answer

---

🚀 Quick Start

1️⃣ Clone the repository

git clone https://github.com/cloudwhj/DataCommAssm3.git

2️⃣ Create and activate a virtual environment

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

3️⃣ Install dependencies

pip install -r requirements.txt

4️⃣ Run the chatbot

streamlit run app.py

The app will start on http://localhost:8501

---

📂 Project Structure

academic-policies-chatbot/
├── app.py                     # Main Streamlit application
├── requirements.txt
├── .gitignore
├── data/
│   ├── policies/              # Policy JSON files
│   └── index/                 # Auto-generated FAISS index
└── README.md

---

🧩 How It Works

1. Policy Data → PDFs are converted into JSON (metadata, structure, clauses, subclauses).
2. Indexing → Text chunks embedded with sentence-transformers and stored in FAISS.
3. Retrieval → Queries are embedded and matched against the index.
4. Generation → Claude (Sonnet/Haiku) answers using the retrieved clauses.
5. Clarification → The chatbot rewrites vague queries (“this”, “it”) and asks clarifying questions when needed.

---

🔒 Responsible-AI Rules

• Responds only using official RMIT policies.
• Adds citations to every factual statement.
• Politely refuses unrelated or personal queries.
• Displays a clear disclaimer for transparency.

---

🧠 Example Queries

Try:

• “What counts as plagiarism?”
• “How do I apply for an assessment extension?”
• “How do I appeal a final course result?”
• “What happens if I fail the same course twice?”

---

⚙️ Default Settings

Parameter	Default	Description
Top-k	6	Number of retrieved clauses
Min similarity	0.35	Confidence threshold
Temperature	0.2	Creativity control
Top-p	0.9	Response diversity

---

🧰 Technologies Used

• Streamlit — Web UI framework
• AWS Bedrock (Claude 3) — LLM inference
• boto3 — AWS SDK for Python
• sentence-transformers — Embeddings
• FAISS — Vector similarity search
• PyPDF2 — PDF parsing

---

🏗 Architecture

 ┌──────────────┐      ┌──────────────┐      ┌──────────────────┐
 │  Streamlit UI │──→──│  Retrieval (FAISS) │──→──│  Bedrock Claude API │
 └──────────────┘      └──────────────┘      └──────────────────┘
        ▲                                         │
        │                                         ▼
        └───── Temporary AWS Credentials via boto3

---

🧪 Testing

Once the app is running, go to:

http://localhost:8501

and test various queries.
Each response includes relevant policy citations for verification.

---

👥 Authors

• Wong Hon Jun – s4060180
• Student 2 – s
• Student 3 – s RMIT University, Bachelor of Computer Science

---

🪪 License

This project is for educational purposes under RMIT’s Academic Use Policy.
Do not redistribute or deploy publicly with institutional credentials.