Enhancing ChatBI System Performance in Vertical Business Domains through an Automated Sample Generation Pipeline

Chat Business Intelligence (ChatBI) systems often fail in vertical business domains due to a critical lack of domainspecific training samples, leading to poor query performance and user trust. This paper addresses this challenge by proposing a novel, automated sample generation pipeline. Our core innovation is a dual-path approach that extracts and synergizes structured knowledge from both internal business documents and database schemas. This allows the system to build a deep, contextual understanding of specialized terminology, metrics, and business logic, even in low-data scenarios. The pipeline then uses this knowledge base to automatically generate a diverse set of high-quality, domain-aware question-and-answer samples. Experimental results on real-world business cases demonstrate the effectiveness of our approach, with NL2SQL accuracy improving dramatically from $34.3 \%$ to $\mathbf{7 1. 5 \%}$. Our work provides a scalable and effective solution for developing reliable and robust ChatBI systems, significantly enhancing their applicability and performance in specialized industries.