PRIME：面向结构化文档检索的预检索索引建模增强框架

View PDF HTML (experimental)

Abstract:High-precision retrieval over external knowledge bases is essential for LLM agents to operate effectively. However, existing Information Retrieval (IR) research primarily optimizes query-side retrievers (e.g., ColBERT, BGE) and generators (e.g., Self-RAG, GraphRAG), while reducing corpus-side representation to chunking. This leaves index-side quality an underexplored lever. We propose Pre-Retrieval Index Modeling & Enhancement (PRIME), an index-side optimization framework built on two complementary tracks: an LLM-driven Semantic Restructuring (SR) track that transforms long documents into structured, information-dense sections, and a deterministic Metadata Synthesis (MS) and Hierarchical Retrieval (HR) track that extracts semantic anchors and heading structure for coarse-to-fine section retrieval. Our evaluations show that deterministic metadata matches structure-aware chunking's retrieval quality while cutting embedding computation by 45.8 $\times$ and index vector count by 22.4 $\times$. On top of these gains, Semantic Restructuring (SR) further improves Answer-F1 by 13.7% in full deployment, with characterized quality-fidelity-cost trade-offs. Across QASPER, BRIGHT, and NFCorpus, we further characterize three qualitative regimes that determine whether index-side optimization helps or degrades retrieval, while showing that these index-side effects are neutralized under a strong reranker.

Comments:	9 pages, 2 figures. Manuscript completed July 2026. Supplementary material will be added in a future version
Subjects:	Information Retrieval (cs.IR); Artificial Intelligence (cs.AI)
Cite as:	arXiv:2610.13296 [cs.IR]
	(or arXiv:2610.13296v1 [cs.IR] for this version)
	https://doi.org/10.48550/arXiv.2610.13296 arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Cheuk Hei Chung [view email]
[v1] Fri, 9 Oct 2026 09:44:47 UTC (2,773 KB)