Year: 2026
Type: article
Abstract: Accessing business information for rapid decision-making today remains significantly hampered by specialized SQL language skills and the problem of hallucination in large-scale language models when working with complex database structures. This research aims to develop a natural language-to-database (NL2SQL) business information querying system and evaluate its technical performance by applying Retrieval-Augmented Generation (RAG) techniques combined with schema injection through the gpt-oss-120b generative AI model on a Groq LPU. The research process involved pilot testing using a standard Spider 1.0 dataset in the department store domain, consisting of 25 questions covering difficulty levels from easy to difficult. The results showed that the developed system performed significantly better than the basic query input method (zero-shot baseline), with an execution accuracy of up to 96.00% and an exact matching (EM) accuracy of 32.00%. Processing time efficiency was also assessed. The system has an average response time of only 0.73 seconds, which meets the human-computer interaction (HCI) benchmark of no more than 3 seconds. The results of this research confirm that integrating database architecture with generative artificial intelligence technology can effectively solve data hallucination problems and reduce communication latency, enabling general users and executives to access insightful data to support business decision-making accurately and quickly at a real-world level.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7216840763