Feedback

Faculté des Sciences appliquées
Faculté des Sciences appliquées
MASTER THESIS

Design and Evaluation of Vector-Based and Graph-Enhanced RAG Architectures for Engineering Knowledge Retrieval

Download
Michel, Merlin ULiège
Promotor(s) : Louppe, Gilles ULiège
Date of defense : 29-Jun-2026/30-Jun-2026 • Permalink : http://hdl.handle.net/2268.2/26123
Details
Title : Design and Evaluation of Vector-Based and Graph-Enhanced RAG Architectures for Engineering Knowledge Retrieval
Translated title : [fr] Conception et évaluation d’architectures RAG fondées sur les vecteurs et enrichies par graphe pour la récupération de connaissances en ingénierie
Author : Michel, Merlin ULiège
Date of defense  : 29-Jun-2026/30-Jun-2026
Advisor(s) : Louppe, Gilles ULiège
Committee's member(s) : Ittoo, Ashwin ULiège
Debruyne, Christophe ULiège
Language : English
Number of pages : 86
Discipline(s) : Engineering, computing & technology > Civil engineering
Institution(s) : Université de Liège, Liège, Belgique
Degree: Master : ingénieur civil en science des données, à finalité spécialisée
Faculty: Master thesis of the Faculté des Sciences appliquées

Abstract

[en] This thesis investigates the use of conversational artificial intelligence to improve access to industrial engineering documentation at Safran Aero Boosters. In this context, engineers and technical teams rely on large volumes of internal documents to retrieve procedures, design knowledge, technical rules, and past experience. However, this information is often difficult to access because it is distributed across heterogeneous documents, written with specialized terminology, and connected through implicit references.

The objective of this work was to design and evaluate a conversational assistant able to provide relevant, reliable, and traceable answers based on private company documentation. To avoid relying only on the internal knowledge of a language model, the system was built around a retrieval-augmented generation pipeline. The pipeline processes the user question, retrieves relevant information from the document base, and generates an answer grounded in the retrieved context.

Two retrieval strategies were compared. The first combines semantic search with keyword-based search in order to retrieve relevant text passages efficiently, including technical terms, acronyms, and document references. The second introduces a graph-based representation of the documentation, aiming to identify technical concepts and relationships between them so that information distributed across several documents can be better connected.

An automatic evaluation was conducted using a synthetic dataset generated from the documentation. The comparison considered answer quality, relevance of retrieved information, grounding, response time, and computational cost. The results show that the graph-based approach generally produced more complete and relevant answers, especially for complex questions requiring several pieces of information. However, it was also slower, more resource-intensive, and sometimes less strictly grounded than the traditional retrieval approach.

The thesis concludes that graph-enhanced retrieval is a promising direction for industrial engineering knowledge access, particularly when contextual richness is more important than speed or cost. Nevertheless, traditional retrieval remains more suitable for faster and more cost-efficient conversational use cases.


File(s)

Document(s)

File
Access MICHEL_Merlin_Master_Thesis.pdf
Description:
Size: 8.44 MB
Format: Adobe PDF

Author

  • Michel, Merlin ULiège Université de Liège > Mast. ing. civ. sc. don. fin. spéc.

Promotor(s)

Committee's member(s)

  • Ittoo, Ashwin ULiège Université de Liège - ULiège > HEC Liège : UER > UER Opérations : Systèmes d'information de gestion
    ORBi View his publications on ORBi
  • Debruyne, Christophe ULiège Université de Liège - ULiège > Dép. d'électric., électron. et informat. (Inst.Montefiore) > Représentation et ingénierie des données
    ORBi View his publications on ORBi








All documents available on MatheO are protected by copyright and subject to the usual rules for fair use.
The University of Liège does not guarantee the scientific quality of these students' works or the accuracy of all the information they contain.