PRIME is an index‑side optimization framework that operates on two complementary tracks. The first track, Semantic Restructuring (SR) driven by large language models, rewrites long documents into structured, information‑dense sections. The second track combines deterministic Metadata Synthesis (MS) with Hierarchical Retrieval (HR) to extract semantic anchors and heading hierarchies for coarse‑to‑fine section retrieval.
Empirical results show that deterministic metadata achieves retrieval quality comparable to structure‑aware chunking while cutting embedding computation by 45.8× and reducing the number of index vectors by 22.4×. On top of these savings, SR further boosts Answer‑F1 by 13.7%. Across the QASPER, BRIGHT, and NFCorpus benchmarks we identify three qualitative regimes that indicate when index‑side optimization helps or harms performance, and we demonstrate that a strong reranker can neutralize these effects.
The key insight of PRIME is to make document structure explicit, allowing the retrieval system to perform an efficient coarse ranking with high‑quality index vectors before query processing, thereby dramatically lowering computational cost and improving accuracy.
Review