Semantic-Type Query Routing for Long-Document Retrieval-Augmented Generation

Abstract

Long-document retrieval-augmented generation (RAG) systems often apply a fixed retrieval strategy to different query types, although fact lookup, summarization, comparison, and multi-hop reasoning require different evidence structures. This paper presents HEC-RAG, a semantic-type routing framework for long-document RAG. HEC-RAG classifies each query into one of four semantic types, assigns an explicit retrieval control operator, performs a section-to-chunk vector similarity cascade, evaluates evidence confidence, and triggers corrective retrieval when evidence is uncertain. Experiments using LongBench and QASPER data show that the LLM Router improves multi-hop recall from 4.0% to 62.19% over rule-based routing. Ablation results show that hierarchical retrieval, semantic routing, and corrective retrieval provide complementary gains.