A Dual-Model Solution for Business Reasoning with Large Language Models

Recent advancements in large language models (LLMs) have shown significant potential in automating tasks using natural language processing. However, as industry demands grow and generative AI advances, more complex tasks—such as intelligent reasoning, strategic decision-making, and business analysis—remain challenging for foundational models trained on general datasets. These tasks require domain-specific knowledge and a deeper contextual understanding, which current models struggle to provide. This research presents an approach that fine-tunes LLMs for specific business use cases by integrating domain knowledge and data analysis capabilities. Leveraging recent developments in structured query (SQL) based fine-tuned models and semantic search techniques, our methodology combines structured data retrieval from organizational sources with relevant domain expertise. This enables the model to answer complex analytical questions by connecting factual data from generated SQL queries to relevant business knowledge. We propose a dual-model system: a fine-tuned code-generating model for data extraction and a foundational model with visual reasoning capabilities. This combination ensures accurate alignment of data with the appropriate knowledge base, reducing the risk of hallucinated answers. The performance of our approach is compared against stand-alone fine-tuned models and retrieval-augmented generation (RAG) models, demonstrating superior results in answering reasoning-based questions and supporting decision-making tasks. Overall, this research offers a systematic solution for integrating data-driven insights with business-specific knowledge, empowering leaders to make informed decisions based on accurate, visualized data and market-specific information.