Design and analysis of an intelligent audit risk assessment system based on retrieval-augmented large language model

Audit risk assessment is increasingly challenged by complex financial environments and the limited ability of traditional data-driven models to capture contextual information. Existing approaches mainly depend on structured financial indicators and often overlook regulatory knowledge and narrative disclosures. This study proposes an intelligent audit risk assessment framework that integrates a retrieval-augmented large language model (RAG-LLM) with a Bayesian–logistic hybrid estimation mechanism. The framework integrates financial, governance, operational, and textual data with audit standards, regulatory cases, and historical accounting irregularities to support context-aware risk evaluation. The proposed system was evaluated using an economically calibrated synthetic dataset consisting of 612 firm-year observations across 14 industries from 2018 to 2024. Results show that the RAG–LLM Bayesian hybrid model achieves an out-of-sample AUC of 0.912, outperforming gradient boosting (0.754), random forest (0.723), and logistic regression (0.694). It also achieves improved calibration performance with a Brier score of 0.088 and a misclassification rate of 6.7%. Further analysis shows that LLM-derived reasoning constructs contribute 48.3% of the total explanatory contribution. The findings demonstrate that retrieval-enhanced reasoning improves prediction accuracy, interpretability, and decision support capability for intelligent auditing systems.