Setio Basuki, Amelia Khoidir, Muhammad Ilham Perdana, Muhammad Daffa Nugraha, Masatoshi Tsuchiya

SC-Literature Intelligence: A Retrieval-Augmented Generation Framework for Multi-Category AI Literature Synthesis in Supply Chain

Introduction

Sc-literature intelligence: a retrieval-augmented generation framework for multi-category ai literature synthesis in supply chain. SC-Literature Intelligence: RAG framework for AI literature synthesis in supply chain. Analyze trends, detect gaps, and synthesize research for structured, evidence-based insights.

Abstract

This paper develops SC-Literature Intelligence, a retrieval-augmented generation (RAG) framework for research synthesis of scientific literature on artificial intelligence (AI) in the supply chain domain. The study addresses the fragmentation of scientific findings, which makes cross-document understanding difficult by supporting four categories of literature-analysis queries: trend analysis, gap detection, comparative synthesis, and evidence-based question answering (QA). The primary novelty lies in introducing a category-aware research synthesis framework capable of evaluating RAG performance across multiple literature-analysis tasks rather than conventional question answering. The framework is built from Scopus-indexed abstracts through pre-processing, chunk-based embedding using BGE-M3 and LaBSE, vector storage, semantic retrieval, and prompt-guided generation evaluated using the RAGAS framework across 640 experimental runs. The results show that BGE-M3 consistently outperforms LaBSE on all RAGAS indicators with the best configuration (chunk size 64, Top-K 5) achieving scores between 0.722 and 0.856 across faithfulness, answer relevancy, context precision, and context recall. Gap detection emerges as the best-supported query category, whereas comparative synthesis remains the most challenging. Failure analysis further reveals that retrieval-stage issues dominate over generation-stage issues, identifying embedding quality as the primary bottleneck. These findings demonstrate that category-aware RAG-based synthesis can support structured, evidence-grounded literature analysis in the supply chain AI domain.

Review

This paper introduces "SC-Literature Intelligence," a timely and relevant Retrieval-Augmented Generation (RAG) framework designed to synthesize scientific literature on Artificial Intelligence within the supply chain domain. The authors effectively address a critical challenge faced by researchers: the fragmentation of findings that hinders comprehensive cross-document understanding. A significant strength lies in its ambition to support a multi-faceted approach to literature analysis, moving beyond conventional question answering to encompass complex tasks such as trend analysis, gap detection, comparative synthesis, and evidence-based QA. This category-aware approach to research synthesis represents a notable advancement over existing RAG applications. The methodological design of SC-Literature Intelligence is robust, detailing its construction from Scopus-indexed abstracts through meticulous pre-processing, chunk-based embedding using BGE-M3 and LaBSE, vector storage, semantic retrieval, and prompt-guided generation. The comprehensive evaluation using the RAGAS framework across 640 experimental runs provides strong empirical evidence for the framework's capabilities. Key findings indicate that BGE-M3 consistently outperforms LaBSE across all RAGAS indicators, with the optimal configuration (chunk size 64, Top-K 5) achieving commendable scores ranging from 0.722 to 0.856. Interestingly, while gap detection emerges as the most effectively supported query category, comparative synthesis presents the most significant challenges, offering valuable insights into the differing complexities of literature analysis tasks. Furthermore, the failure analysis, attributing issues predominantly to the retrieval stage and pinpointing embedding quality as the primary bottleneck, is particularly insightful. Overall, SC-Literature Intelligence demonstrates a promising pathway for automating and enhancing structured, evidence-grounded literature analysis in specialized domains like supply chain AI. The practical implications for researchers are substantial, potentially accelerating knowledge discovery and identifying areas for future investigation with greater efficiency. The identified bottleneck regarding embedding quality provides a clear direction for future research, suggesting that advancements in embedding models or context-aware retrieval strategies could further elevate the framework's performance, especially for challenging tasks like comparative synthesis. This work makes a significant contribution to the field of scientific text mining and RAG applications, offering both a novel framework and critical insights into the performance drivers for complex literature synthesis tasks.

Full Text

Comments

You need to be logged in to post a comment.