Chronological Passage Assembly for Retrieval-Augmented Generation in Narrative Question Answering

Long-context question answering over narrative documents remains challenging because many questions require reconstructing event sequences while preserving local contextual flow under limited context budgets. Existing retrieval-augmented generation (RAG) methods typically retrieve document snippets independently, which can fragment narratives and harm temporal dependencies. We propose ChronoRAG, a retrieval framework for narrative question answering that first converts sequential document chunks into concise relation descriptions and then retrieves relevant units together with their adjacent chronological context. This design preserves retrieval precision while providing the generator with coherent local narrative structure. Experiments on NarrativeQA and GutenQA show that ChronoRAG improves performance on NarrativeQA and remains competitive on GutenQA, with particularly strong gains on questions that require chronology-sensitive context. These results suggest that explicitly modeling local event order is a useful retrieval signal for narrative question answering.