Consultas interativas em banco de dados usando text2SQL baseado em LLM

Abstract

This work presents a systematic analysis of the impact of different prompting techniques and vector distance metrics on the conversion of natural language queries into SQL using large language models (LLMs). The study evaluated the standard, Few-Shot, and Chain of Thought prompting strategies combined with vector similarity metrics such as cosine, Euclidean, dot product, and Manhattan, aiming to identify the most effective combinations for semantic retrieval and SQL query generation. The experiments showed that the Few-Shot technique combined with the cosine distance metric achieved the best results in terms of semantic similarity, correct query execution, and textual quality. To support the analyses, an application was developed based on a RAG architecture, integrating a vector database (Qdrant), the VannaAI library, and the LLaMA 3.1 8B model. This application served as a support tool for the experiments and demonstrates the practical applicability of the analyzed combinations