💡 Project Vision
Developed as a thesis project at the University of Ferrara, this system implements a high-performance semantic document search engine. Unlike traditional keyword-based systems, it understands the intrinsic meaning of queries to provide contextually relevant results.
Conceptual Visualization: Intelligent Document Organization
🎯 Core Objectives
Semantic Indexing
Organizing documents via vector embeddings instead of simple text matching.
Intelligent Retrieval
Finding the most conceptually related documents using Cosine Similarity metrics.
Consumer Hardware
Optimized to run efficiently on mid-range consumer laptops using 4-bit quantization.
🏗️ Technical Architecture
The system is designed with a modular three-tier architecture for maximum portability:
- LLaMa 2 (7B-chat): Meta's next-gen language model used as the core embedding engine.
- Quantization & LLaMa.cpp: Compressing the model from 13GB to 3.9GB to enable local execution on standard CPUs.
- Elasticsearch Vector DB: Utilizing the kNN (k-Nearest Neighbors) algorithm for ultra-fast vector retrieval.
- LangChain: Orchestrating the complex workflow from raw document to searchable vector.
🛠️ Tech Stack
🎓 Thesis Project: Supervised by Prof. Ing. Mauro Tortonesi. This work demonstrates the feasibility of localized, private AI for advanced data management.