Bachelor Thesis — Information Systems Management, HEG Geneva, 2024-2025 · 12 ECTS credits · 360h

SLM & Generative AI: RAG in an Educational Context

Context and Motivation

The rise of generative artificial intelligence is profoundly transforming learning and training practices. This bachelor thesis stems from personal observation and field experience: personalised learning remains difficult to implement at scale, particularly in an academic setting such as HEG, where course materials are numerous, heterogeneous, and hard to leverage in an individualised way.

Rather than relying on costly proprietary Large Language Models (LLMs), dependent on external infrastructure, this work questions the relevance of Small Language Models (SLMs) as a pragmatic alternative, especially when combined with a Retrieval-Augmented Generation (RAG) architecture.

State of the Art

The thesis provides an overview of generative AI and its recent evolution, before laying out the theoretical foundations required to understand the subject: neural network architectures, recurrent neural networks (RNN), and the Transformer architecture, which has become the standard for current language models.

A review of academic and industry literature situates the project relative to existing research on AI in education and personalised learning, as well as to solutions already deployed in the market (Duolingo and other AI-assisted learning platforms).

This analysis leads to the formulation of the research questions, the expected contributions — both technical and pedagogical — and the precise definition of the project scope.

SLM vs LLM: Comparative Analysis

A detailed comparative analysis is conducted between Small Language Models and Large Language Models, covering:

• the classification criteria used to distinguish these two categories of models;
• the advantages and disadvantages of each approach across different use cases (cost, latency, data privacy, infrastructure requirements);
• a comparative benchmark including an analysis of operational costs.

This comparison provides the technical justification for choosing an SLM for the remainder of the project.

RAG Methodology for Personalised Learning

The methodological core of the thesis is the RAG architecture, which enhances a language model's capabilities by giving it access to an external knowledge base at inference time, without requiring a full model retraining.

This section covers:

• the general principles and architecture of RAG (indexing, retrieval, generation);
• the use of vector databases and semantic search techniques;
• the adaptations required to make this architecture compatible with the constraints of an SLM;
• the design choices that orient the system towards effective pedagogical personalisation.

Implementation and Experimentation

A functional prototype is developed using the following technology stack:

The choice of SLM is justified against the criteria defined in the comparative section. The selected use case aims to integrate, as comprehensively as possible, all HEG course materials as a knowledge base, in order to develop a personalised learning assistant prototype applicable in a real academic context.

Evaluation and Results

The prototype is evaluated according to a defined testing protocol and metrics specifically designed for this project. Performance analysis identifies the system's strengths as well as its current limitations, particularly regarding answer accuracy, retrieved context relevance, and scalability.

These results open a discussion on improvement perspectives and pathways towards generalising this approach to other educational contexts.

Conclusion

This bachelor thesis demonstrates the feasibility of a personalised learning solution combining SLM and RAG in a real academic context, while critically examining the trade-offs between performance, cost, and accessibility that are specific to Small Language Models.

---

Keywords: Generative AI, Small Language Models, Retrieval-Augmented Generation (RAG), personalised learning, transformers, vector databases, education

Tech stack: Python · LangChain · ChromaDB · vLLM