Retrieval Granularity as Evidence Design in Small-Model RAG Question Answering: A Diagnostic HotpotQA Study

Abstract

This study treats retrieval granularity as an evidence-design decision in retrieval-augmented question answering. Using HotpotQA, it separates evidence recovery, answer quality, context-budget effects, retriever choice, reranking, runtime and resource use, and subgroup error patterns for small language models.