An Intelligent Recruitment System using Retrieval-Augmented Generation and Cross-Encoder Reranking for Automated Resume Screening
Year: 2026
Type: conference-paper
Abstract: Screening large volumes of unstructured Curriculum Vitae (CV) documents introduces significant information uncertainty: recruiters face difficulty distinguishing relevant candidates from distractors when traditional keyword-based Applicant Tracking Systems (ATS) fail to capture latent semantic relationships. This paper presents an AI-powered Retrieval-Augmented Generation (RAG) framework that systematically reduces retrieval uncertainty through a two-stage hybrid pipeline. In the first stage, a bi-encoder produces dense vector representations for scalable approximate nearest-neighbor search over a repository of 247,480 real-world CV profiles. In the second stage, a cross-encoder reranker jointly processes each query–candidate pair to refine the ranking by capturing fine-grained term interactions that the bi-encoder’s compressed representation inevitably loses. We conduct a controlled five-way ablation study across BM25, bi-encoder, RAG-only, reranker-only, and the full RAG-plus-reranker pipeline, evaluated on Precision@K, NDCG@K, MRR, and query latency. Experimental results demonstrate that the full pipeline achieves a 100.8% improvement in Precision@3 and a 50.0% gain in NDCG@5 over the bi-encoder baseline, while maintaining sub-200 ms end-to-end latency. These findings confirm that cross-encoder reranking serves as an effective uncertainty-reduction mechanism, offering a principled trade-off between retrieval speed and ranking precision for AI-powered human resource technology.
Language: English
Topic: AI and HR Technologies
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7204806428
