How do Text2SQL and language models speed up data access?

A team of researchers representing the WSEI Academy in Lublin, the Netrix S.A. Research and Development Center and the Lublin University of Technology presented an innovative approach to automating database queries, using Text2SQL technology in combination with large language models (LLM). The research results published in the European Research Studies Journal present groundbreaking observations on the effectiveness and practical applications of the Llama3:70b-instruct, Gemma2:27b and Codegemma models in the context of customer service.

The study focused on assessing the precision and speed of the aforementioned language models when generating SQL queries based on instructions expressed in natural language. The Llama3 and Gemma2 models demonstrated high efficiency, achieving correctness of 5 out of 6 queries. In turn, Codegemma, although slightly less accurate, definitely stood out with the shortest response time, which makes it an exceptionally attractive choice for applications where speed of operation is key.

Importantly, the results of the experiment challenged the common belief that providing models with detailed database schemas increases their efficiency. The research team showed that an excess of information in context can lead to reduced efficiency and slower model performance. The need for further development of synonym mapping mechanisms and understanding of industry terminology was also identified, which is essential when handling more complex queries.

The introduction of Text2SQL technology to the customer service environment eliminates the need for staff to know SQL, which significantly simplifies the process of obtaining data from databases, shortening the response time and minimizing the risk of errors. An example of an advanced application of this technology is the Vanna.AI tool, which uses the Retrieval Augmented Generation (RAG) mechanism, combining language models with vector databases and classic database systems. This type of architecture enables smooth and contextually accurate query formulation, increasing the efficiency of interaction with the system.

The conclusions from the study are the foundation for further development of intelligent tools supporting the automation of business processes. The authors recommend simplifying the input context for models, developing synonym recognition systems, and further adapting large language models to the specifics of the query language and domain databases.

The full version of the article is available at:
https://ersj.eu/journal/3498