XRAGLog: A Resource-Efficient and Context-Aware Log-Based Anomaly Detection Method Using Retrieval-Augmented Generation

Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Hongyi Liu, Ying Li

Keywords: Anomaly Detection, Log Compression, Retrieval-Augmented Generation

Abstract: As large language models (LLMs) become more sophisticated and pervasive, ensuring their operational stability has become increasingly critical. Consequently, the need for accurate and reliable anomaly detection has grown significantly. Leveraging the rich semantic information within log data, LLMs have proven to be powerful tools for anomaly detection. However, existing log-based anomaly detection methods that utilize LLMs are resource-intensive and often overlook the interrelationships between log entries. To address these limitations, we propose a resource-efficient and context-aware log-based anomaly detection approach. This method combines hierarchical log compression with context-aware retrieval-augmented generation to enhance efficiency and accuracy. Experiments on various public and real-world datasets demonstrate that our approach significantly improves anomaly detection accuracy while dramatically reducing token consumption.

Submission Number: 14