A synergistic multi-stage RAG architecture for boosting context relevance in data science literature

Official URL: https://dx.doi.org/10.1016/j.nlp.2025.100179

Navigating the voluminous and rapidly evolving data science literature presents a significant bottleneck for researchers and practitioners. Standard Retrieval-Augmented Generation (RAG) systems often struggle with retrieving precisely relevant context from this dense academic corpus. This paper introduces a synergistic multi-stage RAG architecture specifically tailored to overcome these challenges. Our approach integrates structured document parsing (GROBID), domain-specific embedding fine-tuning derived from textbooks, semantic chunking for coherence, and proposes a novel ’Abstract First’ retrieval strategy that prioritizes concise, high-signal summaries. Through rigorous evaluation using the RAGAS framework and a custom data science query set, we demonstrate that this integrated architecture significantly boosts Context Relevance by over 15-fold compared to baseline RAG, surpassing configurations using only subsets of these enhancements. These findings underscore the critical importance of multi-stage optimization and highlight the surprising efficacy of the abstract-centric retrieval method for specialized academic domains, offering a validated pathway to more effective literature navigation in data science.