Bridging the Chunk Gap: Pre-Computed Narrative Summaries for Temporally Grounded Retrieval-Augmented Generation

Abstract

Retrieval-Augmented Generation (RAG) systems that rely solely on fixed-length text chunksfor retrieval systematically lose the narrativearc and chronological context inherent in temporally structured data sources such as issuetracking records. We propose augmenting theretrieval index with pre-computed NarrativeAnalysis Summaries (NAS)—concise, chronologically faithful synopses generated for eachissue during the offline ingestion phase. Atquery time, a dual-source retrieval mechanismfetches both localised text chunks and the mostrelevant summaries, merges them via Reciprocal Rank Fusion, and applies a cross-encoderreranker to refine the final ranking before generation. Evaluated on a dataset of 500 queriesspanning fact-seeking, historical, and relationalcategories over both proprietary enterprise Jiradata and the public Jira dataset of Montgomeryet al. (2022), our approach yields a 30.4% improvement in Context Recall and a 20.1% improvement in NDCG@5 relative to a chunkonly baseline, while incurring only a 1.7% decrease in Context Precision and achieving a3.0% gain in Faithfulness. These results demonstrate that incorporating pre-computed chronological summaries with cross-encoder reranking substantially enhances retrieval completeness and ranking quality without compromisinggeneration fidelity