A Hybrid Retrieval and Agent Reasoning Approach to KEJI Index Evaluation Using Hybrid-RAG

Due to the increasing importance of ESG (Environmental, Social, and Governance) criteria, the Korea Economic Justice Institute Index (KEJI), a standard framework for assessing social responsibility of Korean companies, has required more accurate and efficient calculations. However, the current KEJI evaluation workflow relies heavily on manually verifying expert analyst annual reports and regulatory disclosures. This process is creating significant labor costs and lack of scalability. In this paper, we propose an automatic extraction method of KEJI indicators based on hybrid search augmented generation (Hybrid RAG). The proposed KEJI indicator automatic extraction method has the steps of (i) multimodal PDF parsing through PyMuPDF, (ii) a hybrid search machine that combines high-density vector embedding with sparse keyword search, (ii) a reciprocal rank fusion (RRF) for optimizing search ranking, and (iv) a React agent-based inference loop to derive iterative evidence and extraction values. In this experiment, a real dataset curated manually was used to evaluate the extraction precision of important indicators, especially total issued stocks, internal and family ownership, and controlling shareholders’ shares. From the experimental results, we confirmed that the standardized table indicators were extracted with high accuracy, but the performance was degraded when parsing irregular table structures into linear text. Through this, the proposed method will provide the effectiveness of RAGbased automation in KEJI analysis and the technical criteria for the automated ESG evaluation system.