SES-RAG: An Adaptive Framework with ModernBERT Segmentation and Generative-Alignment Expansion

Retrieval-augmented generation (RAG) is promising for long-document question answering, but its performance is often limited by fixed-window chunking, query-document mismatch, and static context truncation. This study proposes SES-RAG, a lightweight framework that jointly improves semantic segmentation, retrieval alignment, and context selection in long-document RAG. Specifically, SES-RAG includes a ModernBERT-based semantic segmenter for constructing semantically coherent chunks, Generative-Alignment Query Expansion (GAQE) for enhancing query-document alignment through residual fusion with a hypothetical document, and Information Density Truncation (IDT) for adaptively selecting retrieved contexts. We evaluate SES-RAG on three long-document QA benchmarks, namely NarrativeQA, QuALITY, and QASPER, against Naive RAG, HyDE-RAG, Hybrid-RAG, and CRAG. SES-RAG achieves the best overall performance on all three datasets, reaching 15.71 ROUGE, 9.61 BLEU-1, 6.16 BLEU-4, and 18.71 METEOR on NarrativeQA, 74.33 accuracy on QuALITY, and 41.06 F1-Match on QASPER. Compared with Naive RAG, SES-RAG improves ROUGE by 5.10 points on NarrativeQA, Accuracy by 3.66 points on QuALITY, and F1-Match by 3.83 points on QASPER. These results show that lightweight cross-stage optimization is an effective and practical strategy for improving long-document RAG.