The Impact of Fine-Tuning in Embedding Models on Query-Dependent Retrieval Performance

In the field of Machine Learning, the ability of Large Language Models to generate context-dependent responses is of significant importance in the development of models. Source texts are divided into segments from which embedding vectors are extracted and associated with incoming prompts to select contexts relevant to the request. This process not only increases the accuracy of responses but also prevents the risk of hallucinations. Therefore, the methods used to calculate vector similarities between context and prompt have a notable impact on model performance. This study focuses on the role of embedding vectors in retrieving information relevant to a prompt. A fine-tuning process performed to improve the efficacy of these vectors in extracting information. The impact of the fine-tuning process on a domain-specific dataset is examined within this scope. Additionally, the effectiveness of semantic similarity functions used in the context retrieval process according to the prompt is evaluated for their success in finding similarities.