Schema Retrieval with Embeddings and Vector Stores Using Retrieval-Augmented Generation and LLM-Based SQL Query Generation

In today’s world, where the volume and variety of data are increasing at an extraordinary rate, extracting meaningful insights from data is of critical importance; however, the complexity of standard database query languages makes it difficult for users without technical expertise to access information. This study proposes an innovative Retrieval-Augmented Generation (RAG) architecture that analyzes natural language queries, identifies related database schemas, and automatically converts them to SQL. Unlike fixed schema selection (fixed-k) methods, a unique hierarchical clustering mechanism is introduced to dynamically determine the number of relevant schemas, minimizing noise. Furthermore, the architecture incorporates an iterative repair mechanism, data enrichment with sample rows, and a hybrid query strategy (Turkish + English) to overcome cross-lingual barriers. Performance evaluations on 15 databases demonstrate that the proposed method improved the schema retrieval F1 score from 0.79 to 0.88. In the SQL generation phase, the execution accuracy (EX) of the GPT-4o model increased from 0.70 to 0.78 with the proposed optimizations, representing an approximate 11% improvement relative to the baseline configuration without requiring fine-tuning.