LLM-Supported Hybrid Reranking Architecture for Candidate-Job Matching in Recruitment Processes
TL;DR
This study proposes a hybrid architecture that integrates vector-based semantic search with data validation steps and delivers the final decision through a large language model-based reranking module to overcome problems with traditional applicant tracking systems and pure large language model methods.
Abstract
Traditional applicant tracking systems miss candidates' potential due to their keyword-based structures, while pure large language model approaches carry high costs and hallucination risks. This study proposes a hybrid architecture that integrates vector-based semantic search with data validation steps and delivers the final decision through a large language model-based reranking module to overcome these issues. The proposed system processes candidate data through a multi-stage filtering funnel by separating structured and unstructured formats. Experimental analysis on a real-world dataset comprising 52 job postings and 36 candidate resumes demonstrates that this hybrid approach achieves a retrieval success rate of 91.67%, a human resources expert scoring alignment of 86.11%, and a mean absolute error of 0.200, significantly outperforming traditional applicant tracking systems and pure large language model methods in both accuracy and cost-efficiency. The findings show that the architecture achieves high retrieval success and scoring alignment, offering an explainable, scalable, and accurate decision support system for human resources processes. Ultimately, this study bridges the gap between theoretical RAG capabilities and practical human resources deployment, presenting a reliable framework for next-generation recruitment.