Large language models for automating database operations in relational databases

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Controlling database systems using traditional interfaces requires a strong understanding of Structured Query Language (SQL) and the database administration tools employed. This study introduces a new AI-based system that automates create, read, update, and delete database actions through natural language processing, supporting multiple natural languages and diverse user vocabularies. The fundamental architectural innovation of this research consists of developing a bridge between Google's Gemini Large Language Model and the Google Agent Development Kit, referred to as the Gemini-Agent Development Kit Bridge. While the Agent Development Kit is built on a set of structured commands and handles all database query execution, Google's Gemini serves as a smart communication layer between the user and the Agent Development Kit. Thus, the main responsibility of Gemini is the translation from the entire range of natural language used by the user into a structured command recognizable by the Agent Development Kit, regardless of its vocabulary, informality, or language. In the same way, the output of the Agent Development Kit is transformed into human-readable output by Gemini. The proposed solution was designed as a Flask-based e-commerce management system and evaluated using 325 manually authored test cases covering five operational classes. On average, 93.8% accuracy was achieved for the translation of commands, with an accuracy of 96.0% for order deletion and 87.5% for complex multi-step operations. The response time varied between 1.2 seconds for simple commands and 3.5 seconds for complex operations. This finding has substantial implications, as it suggests that Gemini's broad natural-language capabilities can be employed to expand access to structured backend systems without altering the execution layer.