🏛️ VGU One Ai Assistant (RAG)

https://ragbotjarvis-twhygdejckztzwcyyyrypu.streamlit.app/

An intelligent, authoritative student success AI designed to provide accurate information from the VGU Student Handbook, Hostel Booklet, Club Brochure and ACIC Document. This project demonstrates a production-ready Retrieval-Augmented Generation (RAG) pipeline.

---

🛠️ Infrastructure & Tech Stack

🚀 Backend: Hugging Face Docker Space (n8n)

The "Brain" of this project is a self-hosted n8n orchestration engine running within a custom Docker container on Hugging Face Spaces.

🐧 Backend Environment (n8n + Docker)

The orchestration layer is hosted in a Hugging Face Docker Space using a customized node:20-alpine image.

• Alpine Linux Base: I selected a minimal Alpine-based image to reduce the container footprint, resulting in faster deployment cycles and lower memory overhead.
• Node.js 20: Ensures compatibility with the latest n8n enterprise features and AI nodes.

💾 Data Persistence & Volume Mapping

Hugging Face Spaces utilize ephemeral storage, meaning any data saved locally is lost upon container restart. To solve this:

• Custom Volume Mapping: I implemented a persistent storage strategy by mapping the n8n user folder (/root/.n8n) to a dedicated volume.
• Persistence: This ensures that n8n credentials (Gemini/Supabase API keys) and workflow definitions remain intact even if the cloud instance reboots.

🛡️ Security & Access Control

Deploying an AI to the public web requires specific security hardening:

• Permission Overrides: To bypass the strict "Read-Only" filesystem of the HF environment, I applied specific chmod 777 permission overrides during the build process to allow n8n to initialize its internal SQLite database successfully.
• RAG-Based Hardening: The system is protected against "Prompt Injection" by a strict system prompt that forces the LLM to derive answers EXCLUSIVELY from the Student Handbook.
• Endpoint Obscurity: The n8n backend is protected via a unique UUID-based webhook endpoint, preventing unauthorized API calls from external sources.

🧠 Intelligence: Google Gemini 2.5 Flash Lite

• Model: gemini-2.5-flash-lite
• Role: Synthesis of retrieved handbook text into natural, professional language.
• Constraint: Strictly limited to provided context to ensure zero hallucinations.

📁 Database: Supabase (pgvector)

• Storage: Vectorized embeddings of the documents in SUPABASE.
• Retrieval: Similarity search using text-embedding-001 to match user queries with handbook clauses.

🎨 Frontend: Streamlit

• UI: Custom-styled to match the official VGU One Portal.
• UX: Top-navigation tabs for seamless switching between Chat, Architecture, and Contact.

---

🏗️ System Flow

1. Request: User asks a query on the Streamlit frontend.
2. Trigger: Streamlit sends a POST request to the n8n Production Webhook on Hugging Face.
3. Search: n8n queries Supabase via vector similarity search.
4. Prompting: n8n sends the retrieved "Context" + User "Query" to Gemini Flash.
5. Output: The LLM returns the answer with a PDF Page Citation, which is then rendered on the UI.

---

📖 Installation & Usage

Backend (n8n on Hugging Face)

1. Use the provided Dockerfile "docker-hugging-space.txt" to spin up an n8n instance on HF Spaces.
2. Set Environment Variables: N8N_PORT=7860, DB_TYPE=sqlite, N8N_USER_FOLDER=/root/.n8n.
   • 🛠️ One quick warning on URLs Hugging Face can be picky about trailing slashes. When you fill in your variables, make sure they look like this: WEBHOOK_URL: https://<user_name>-<space_name>.hf.space (No / at the end) N8N_HOST: <user_name>-<space_name>.hf.space (No https:// and no /)
3. Import VGU_Workflow.json.

Frontend (Streamlit)

1. Clone this repo.
2. Update N8N_WEBHOOK_URL in app.py with your HF Space link.
3. Run streamlit run app.py.

---