Custom Data Integration for Large Language Models: Investigating Fine-tuning and Retrieval-Augmented Generation

Natural Language Processing (NLP) has witnessed remarkable advancements in recent years, enabling machines to comprehend and generate human-like language. In this context, my master’s thesis aims to design and implement a semi-automatic process that allows language models to interact with custom datasets. The project specifically focuses on enabling language models to answer questions related to given tabular and text-based data. To achieve this goal, two main approaches were explored, the first one involves leveraging a Large Language Model (LM) to generate data descriptions and question-answer pairs in order to fine-tune a target LLM and increase the model’s knowledge of a specific topic. This method is more expensive and it is technically more difficult to carry out. The second approach deals with Retrieval Augmented Generation (RAG), in which data descriptions are transformed into embeddings, representing their semantic relationships. These embeddings are then used to retrieve the most relevant texts given a query and to integrate them into the prompt presented to the LLM during inference. It makes it possible to avoid the costs and knowledge required to fine-tune a model and allows to use a language model by entering the most relevant texts into the prompt with the original query.