RAGtor-UPV: AI Chatbot for UPV Regulations

Project Overview

RAGtor-UPV is an AI-powered chatbot designed to assist students, staff, and faculty at the Universitat Politècnica de València (UPV) by answering questions about university regulations, procedures, and academic policies. The system leverages Retrieval-Augmented Generation (RAG) with a custom vector database (FAISS) and a large language model (PoliGPT) to provide accurate, context-aware, and up-to-date responses based on official university documents.

---

Features

• Conversational Chatbot: Friendly, interactive chat interface built with Streamlit.
• Contextual Retrieval: Uses semantic search (FAISS + sentence-transformers) to find the most relevant regulatory content.
• RAG Pipeline: Combines retrieved context with a generative LLM (PoliGPT) for precise, grounded answers.
• PDF & Database Management: Automatically scans, indexes, and updates regulatory PDFs in a SQLite database.
• Parallel Processing: Efficient PDF chunking and embedding using multiprocessing.
• Session Memory: Maintains chat history for context continuity.
• Customizable & Extensible: Modular codebase for easy adaptation to other institutions or document sets.

---

How It Works

1. PDF Ingestion & Database Update

   • Place all regulatory PDFs in the 01_data/pdf_actuales/ directory.
   • Run the database update script to scan, register, and track new or modified PDFs in project_database.db.

2. Chunking & Embedding

   • PDFs are split into overlapping text chunks.
   • Each chunk is embedded using a multilingual transformer model (sentence-transformers/LaBSE).

3. FAISS Vector Store

   • All embeddings are stored in a FAISS index (01_data/project_faiss/) for fast semantic search.

4. Chatbot Interaction

   • Users interact via a Streamlit web interface.
   • When a question is asked, the system retrieves the top-k most relevant chunks from FAISS.
   • The retrieved context and the user’s question are sent to PoliGPT (hosted LLM API).
   • PoliGPT generates a concise, context-grounded answer, citing the relevant regulation.

5. Session & Logging

   • All interactions are logged for future improvements and analytics.

---

Setup Instructions

1. Clone the Repository

git clone https://github.com/your-org/ragtor-upv.git
cd ragtor-upv

2. Prepare the Environment

• Install Python 3.9+ (recommended: use a virtual environment).
• Install dependencies:

python -m venv rag_fix
source rag_fix/Scripts/activate  # On Windows: rag_fix\Scripts\activate
pip install -r requirements.txt

3. Prepare Data

• Place all regulatory PDFs in 01_data/pdf_actuales/.
• (Optional) Place additional CSVs or data in 01_data/DB_CSV/.

4. Initialize the Database

python 01_main/database_sql.py

This will create and populate 01_data/project_database.db with metadata about your PDFs.

5. Build or Update the FAISS Index

python 01_main/database_faiss_murta.py

This script will:

• Chunk and embed all active PDFs.
• Store embeddings in the FAISS index (01_data/project_faiss/).

6. Launch the Chatbot Interface

streamlit run 01_main/interfaz_00_conexion_prueba.py

Open the provided local URL in your browser to start chatting with RAGtor-UPV.

---

Project Structure

01_data/
    pdf_actuales/         # Regulatory PDFs
    project_database.db   # SQLite DB with PDF and chunk metadata
    project_faiss/        # FAISS vector index
    DB_CSV/               # Additional data (optional)
01_main/
    database_sql.py       # DB creation and update scripts
    database_faiss_murta.py # FAISS index management
    poli_gpt.py           # RAG pipeline and PoliGPT API client
    interfaz_00_conexion_prueba.py # Streamlit chatbot UI
static/                   # Static files for web interface
rag_fix/                  # Python virtual environment

---

Typical Workflow

1. Add or update PDFs in 01_data/pdf_actuales/.
2. Update the database:
   python 01_main/database_sql.py
3. Update the FAISS index:
   python 01_main/database_faiss_murta.py
4. Launch the chatbot:
   streamlit run 01_main/interfaz_00_conexion_prueba.py
5. Ask questions about UPV regulations and receive grounded, referenced answers.

---

Troubleshooting

• No vectors in FAISS: Make sure you have run both the database and FAISS update scripts after adding PDFs.
• Timeouts with PoliGPT: Check your internet connection and the availability of the PoliGPT API.
• Database errors: Ensure the database is created and up-to-date before running the FAISS script.
• PDF parsing errors: Some PDFs may be encrypted or malformed; check logs for details.

---

Contributors

• Marxx01 - Marc Hurtado Beneyto
• Hervaas8 - Alejandro Hervás Castillo
• Vimapo23 - Víctor Mánez Poveda
• QuicoCaballer - Francisco Caballer Gutierrez
• ogarmar - Óscar García Martínez

---

License

RAGtor-UPV, Universitat Politècnica de València

---

Acknowledgements

• LangChain
• FAISS
• Streamlit
• sentence-transformers
• OpenAI