Question-Answering System for Geoportale Nazionale Archeologia (GNA)

This repository contains the codebase and evaluation material for the design and development of a Question-Answering (QA) system for the Geoportale Nazionale per l’Archeologia (GNA). The project was carried out as part of the internship in preparation for my Master’s thesis in Digital Humanities and Digital Knowledge at the University of Bologna.

Project Overview

The GNA QA system is a retrieval-augmented question-answering assistant designed to respond to natural language queries based on official GNA documentation. It integrates web crawling, document chunking, vector-based retrieval and generation in a modular and scalable architecture.

The application is available at:

👉 gna-assistant-ai.streamlit.app

Main features include:

• Focused crawling and sitemap generation from GNA wiki operative manual
• Chunked document processing and metadata annotation
• Dense embeddings generation using multilingual Sentence Transformers
• Retrieval-augmented generation
• Citation-aware prompting
• Streamlit-based user interface and feedback tracking
• Evaluation suite for retrieval metrics (Precision@k, Recall@k, MRR)
  
  

📂 Structure

• generate_sitemap.py, create_chunks.py, vector_store.py: Knowledge base preparation
• rag_sys.py: Retrieval-Augmented Generation pipeline
• main.py: Streamlit application logic
• feedback_handling.py: Feedback management
• evaluate_retrieval.py: Evaluation framework
• create_test_data.py: Test set generation
• main_preprocess.py: Combined pipeline for sitemap, chunking, and vectorization
• data/: Document chunks, test datasets, metrics, logs
• feedback/: Local SQLite database for user feedback
• sitemap/: XML sitemap of the GNA website
• OCR/: OCR-related scripts
• .faiss_db/: FAISS vector store
• .streamlit/: Streamlit configuration files
• requirements.txt: Python dependencies
• packages.txt: Additional system requirements for Streamlit Cloud
  
  

📊 Evaluation

Automated evaluation was performed using a synthetic test set of 400 domain-specific questions. Key metrics include:

• Precision@5
• Recall@5
• MRR (Mean-Reciprocal Rank)
• Avg. Retrieval Time

Acknowledgments

This project was supervised and supported by Mario Caruso and Simone Persiani from BUP Solutions, whose guidance and technical insights were instrumental throughout the internship. I sincerely thank them for their time, encouragement, and valuable mentorship.