Λογότυπο αποθετηρίου

Πτυχιακές εργασίες

Μόνιμο URI για αυτήν τη συλλογήhttps://pyxida.aueb.gr/handle/123456789/11719

Περιήγηση

Πλοήγηση Πτυχιακές εργασίες ανά Συγγραφέα "Μητσάκης, Νικόλαος"

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω
Φιλτράρετε τα αποτελέσματα πληκτρολογώντας τα πρώτα γράμματα
Τώρα δείχνει 1 - 1 από 1
  • Δεν υπάρχει διαθέσιμη μικρογραφία
    Τεκμήριο
    Leveraging retrieval-augmented generation for student support: a document-centric QA system for the AUEB informatics studies guide
    (2025-07) Mitsakis, Nikos; Μητσάκης, Νικόλαος; Androutsopoulos, Ion; Stafylakis, Themos
    This thesis examines the design, development, and evaluation of a Retrieval-Augmented Generation (RAG) system specifically designed to support undergraduate students in the Department of Informatics at the Athens University of Economics and Business (AUEB). The central objective is to create a cost-effective yet high-quality AI assistant capable of answering studies guide-related questions, ensuring that all responses are explicitly grounded in the latest edition of the department's official Studies Guide. To achieve this, the system ingests the newest version of the Studies Guide. It represents its contents at three levels of granularity: chunks (bodies of text corresponding to paragraphs or groups of paragraphs on a specific topic, based on the document’s structure), sentences (extracted by sentence tokenizing each chunk), and propositions (decontextualized factual statements synthetically generated from the chunks). The retrieval architecture explores traditional lexical search (BM25), dense vector search, and a hybrid ensemble retriever to maximize retrieval coverage and relevance. Question-answering capabilities are assessed using both real-world and synthetic QA pairs, with the generation module leveraging self-hosted state-of-the-art large language models (LLMs). The thesis conducts a comprehensive evaluation across all document granularities and retrieval configurations, employing both classical information retrieval metrics and more modern LLM-based evaluation. Results demonstrate the feasibility of delivering a factual, responsive, and modular assistant using modest computational resources. The thesis further discusses the limitations and potential extensions of the approach, aiming to provide a blueprint for deploying similar RAG-based assistants in other academic settings.