Perancangan Dan Implementasi Sistem Two-Stage Retrieval Berbasis Bm25 Dan Neural Reranking Pada Dataset Squad 1.1

Abstract

This study designs and implements a two-stage retrieval system combining BM25-based candidate generation with neural reranking, applied to the SQuAD 1.1 dataset reframed for passage retrieval. Development covers data preparation, corpus construction through selective chunking, BM25 candidate generation, and integration of a modular reranking component into an interactive user interface. System evaluation was conducted through black-box testing across multiple functional scenarios. Result show that the corpus and relevance mapping were successfully built, BM25 achieved high candidate recall, and all interface features functioned as designed. These findings confirm that the proposed pipeline can be implemented end-to-end into a functional application, providing a validated basis for further quantitave evaluation of reranking performance.