GPT-3.5 is a language model trained to perform various natural language processing tasks. This capability can be utilized in the low-code platform for web-based applications data management called Qore-Base, by assisting users in setting up database schemas and data authorization settings. Qore-Base provides an API where relevant operations can be written as payloads. The API payload can be obtained from the GPT-3.5 model by providing an input of the application description the user intends to develop. However, the Qore-Base payload is quite complex, while the usage of GPT-3.5 is limited in terms of token count. In this research, a Domain-Specific Language (DSL) has been developed to represent creating database schemas and setting data authorization in a more straightforward language format that can be converted into Qore-Base API payloads. A tool with a user interface has also been built to input the application description, which is the GPT-3.5 model input. Furthermore, experiments have been conducted to obtain the optimal combination of GPT-3.5 model parameters. The testing was performed by comparing the syntactic and semantic scores of the generated DSL for each parameter combination when tested against five queries with five levels of detail and five queries with varying levels of complexity. From the conducted experiments, the best utilization of the GPT-3.5 model was achieved using the base Text-Davinci model with a prompt approach consisting of a brief DSL documentation form accompanied by example query pairs and the expected query answers. Overall, the performance of the DSL for database schemas was nearly excellent, particularly for highly detailed queries. However, the performance of the DSL for data authorization settings needs improvement.
No takes yet. Share an insight, caveat, or question.
Ufairah et al. (2023) studied this question.