LLM-RAG for Financial Question Answering: A Case Study from SET50

This study compares the performance of a traditional relational database with a financial knowledge graph in retrieval-augmented generation (RAG) settings. The knowledge graph contains stock closing prices and financial statement data from companies in the SET50 index, focusing on key financial metrics and market performance indicators. The study examines two main aspects: response time and the accuracy of large language models (LLMs) in query generation. To evaluate the efficiency of both databases, complex queries-such as cross-industry financial ratio comparisons and trend analysis over time-are utilized. The experimental results indicate that the graph database requires less time to retrieve results. Further-more, the LLM can translate natural language into queries for both the graph and relational databases with a similar level of accuracy.