Vector Similarity-Assisted ChatGPT Approach for Text Classification

This study presents a comparative analysis of the effectiveness of large language models and vector similarity-based search methods in text classification tasks. Two distinct datasets are used to evaluate semantic similarity via the FAISS (Facebook AI Similarity Search) library and contextual classification performance through ChatGpt models. Classification was performed using five different FAISS models with varying sizes and architectures. Subsequently, the most semantically similar samples were passed to the ChatGpt API, forming a hybrid approach. The results show that FAISS is particularly effective in small and imbalanced datasets, while ChatGPT-supported architectures significantly improve accuracy in semantically complex content. The impact of preprocessing steps and class imbalance on model performance is discussed, highlighting the importance of dataset characteristics in model selection. The proposed hybrid method is evaluated as an efficient and scalable alternative for applications requiring high accuracy with limited labeled data.