Retrieval-Augmented Generation Strategies for Library Knowledge Services: A Stage-Wise Study

Abstract

Library knowledge services increasingly use retrieval-augmented generation (RAG) for questions that require evidence from multiple documents. Yet RAG pipelines combine retrieval, fusion, reranking, evidence packing, and generation, and comparisons often change several stages together, obscuring where gains arise. Using multi-hop question answering as a controlled proxy, we conduct a stage-wise study that fixes the corpus, questions, generator, prompt, and decoding while varying one stage-level intervention at a time. We compare sparse and dense retrieval, reciprocal-rank fusion, cross-encoder reranking, evidence budgets, and evidence controls using per-question records, paired tests, and stage-level latency. On a frozen HotpotQA-derived corpus, dense retrieval improves candidate discovery, fusion broadens deeper support coverage, and reranking most consistently moves complete title-level support to prompt-relevant ranks and improves answers. Longer contexts increase document coverage and cost but can reduce accuracy, while gold supporting sentences improve accuracy yet leave substantial residual error. The findings offer stage-specific insights for designing and assessing library-oriented RAG pipelines.