Paper deep dive
PRIME: Pre-Retrieval Index Modeling Enhancement for Structured Document Retrieval-Real
Cheuk Hei Chung, Li Liu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 10/12/2026, 7:10:23 AM
Summary
The paper introduces PRIME, an index-side optimization framework for structured document retrieval that enhances information density and structure without modifying the retriever. It comprises two tracks: Semantic Restructuring (SR), which uses LLMs to compress and explicate document sections, and Metadata Synthesis (MS) with Hierarchical Retrieval (HR), which extracts deterministic metadata for efficient filtering and reranking. PRIME demonstrates significant efficiency gains (45.8x reduction in embedding computation, 22.4x reduction in index vectors) while matching or improving retrieval quality over structure-aware chunking, particularly in dense-only pipelines. The study identifies three corpus regimes (deficit-bridging, structural exploitation, structural saturation) and notes that strong cross-encoder rerankers can neutralize index-side optimization benefits.
Entities (13)
Relation Signals (12)
PRIME → evaluatedon → NFCorpus
confidence 98% · Across QASPER, BRIGHT, and NFCorpus, we further characterize three qualitative regimes
PRIME → evaluatedon → BRIGHT
confidence 98% · Across QASPER, BRIGHT, and NFCorpus, we further characterize three qualitative regimes
PRIME → evaluatedon → QASPER
confidence 98% · Across QASPER, BRIGHT, and NFCorpus, we further characterize three qualitative regimes
PRIME → hascomponent → Semantic Restructuring
confidence 95% · PRIME... built on two complementary tracks: an LLM-driven Semantic Restructuring (SR) track...
PRIME → hascomponent → Metadata Synthesis
confidence 95% · PRIME... built on two complementary tracks: ... a deterministic Metadata Synthesis (MS) ... track
PRIME → hascomponent → Hierarchical Retrieval
confidence 95% · PRIME... built on two complementary tracks: ... Metadata Synthesis (MS) and Hierarchical Retrieval (HR) track
Semantic Restructuring → improvesmetric → answer F1
confidence 90% · Semantic Restructuring (SR) further improves Answer-F1 by 13.7% in full deployment
Metadata Synthesis → reducescomputation → embedding_computation
confidence 90% · deterministic metadata matches structure-aware chunking’s retrieval quality while cutting embedding computation by 45.8×
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2610.13296v1
- Canonical: https://arxiv.org/abs/2610.13296v1
Trouble viewing inline? Open PDF directly →
Full Text
45,644 characters extracted from source content.
Expand or collapse full text
PRIME: Pre-Retrieval Index Modeling & Enhancement for Structured Document Retrieval Cheuk Hei Chung Li Liu †thanks: Email: chchung634@connect.hkust-gz.edu.cn†thanks: Corresponding author: Li Liu (avrillliu@hkust-gz.edu.cn) 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× and index vector count by 22.4×. 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. The Hong Kong University of Science and Technology (Guangzhou) 1 Introduction Efficient information retrieval (IR) is essential for LLM agents, as the quality of retrieved contexts serves as the foundational bottleneck for downstream generation (Guu et al. 2020; Lewis et al. 2020; Borgeaud et al. 2022; Izacard et al. 2023). While query-side retrievers and generators have been extensively refined—from sparse models (Robertson and Zaragoza 2009) to dense dual-encoders (Karpukhin et al. 2020), late-interaction frameworks (Khattab and Zaharia 2020), and cross-encoder rerankers—corpus-side representation has largely remained a fixed input. Conventional preprocessing merely alters text without enhancing its information density or logical structure. Figure 1: Overview of Index-Side Optimization. PRIME operates on the corpus representation prior to indexing, complementing existing query-side and generator-side methods without modifying the retriever. Figure 2: PRIME architecture. The framework operates on two complementary tracks: (Left) an LLM-driven Semantic Restructuring track transforms documents into knowledge-dense representations; (Right) a deterministic track uses Metadata Synthesis to extract structural anchors, which power a Hierarchical Retrieval pipeline with RRF fusion and cross-encoder reranking. Recent methods began to exploit document structure for indexing, and Singh et al. (2026) show that once a strong cross-encoder reranker is present, common retrieval enhancements yield no reliable additional gain. This raises the question of whether index-side representation can be improved and whether its benefits are compensated by a strong pipeline. We propose that retrieval quality is bounded by the information density of the indexed corpus. Once a strong retriever and reranker are in place, the remaining leverage lies in information structure and density at the index level. Guided by this principle, we present PRIME, an index-side quality optimization framework that operates on two complementary tracks without modifying the retriever. Semantic Restructuring (SR) employs an LLM-driven transformation to explicate implicit argument structure and compress descriptive prose within each section, increasing knowledge density for long structured documents. In parallel, Metadata Synthesis (MS) and Hierarchical Retrieval (HR) form a fully deterministic track: MS extracts semantic anchors and heading structure from the original document, which HR then consumes with lightweight high-density embeddings for coarse document filtering and a cross-encoder for fine-grained section reranking. Figure 1 situates PRIME in the retrieval landscape. Our key contributions are: • Deterministic structural metadata matches structure-aware chunking’s retrieval quality while cutting embedding computation by 45.8× and index vector count by 22.4×, establishing a more efficient and lightweight alternative to LLM-mediated indexing. • On top of these deterministic gains, Semantic Restructuring (SR) further yields an additional +13.7% Answer-F1 in full deployment, with characterized semantic quality, fidelity, and cost trade-offs. • Through a 5 × 3 factorial ablation on QASPER, BRIGHT, and NFCorpus, we identify three corpus regimes,deficit-bridging, structural exploitation, and structural saturation,that determine whether index-side optimization helps or degrades retrieval, and show that a strong cross-encoder reranker neutralizes these gains. 2 Related Works Query-side retrieval has been extensively optimized. From sparse retrieval (Robertson and Zaragoza 2009) to learned sparse representations such as SPLADE (Formal et al. 2021). Dense retrieval progressed through dual-encoder architectures including DPR (Karpukhin et al. 2020), ANCE (Xiong et al. 2021) and Contriever (Izacard et al. 2021), multi-vector late-interaction models represented by ColBERT (Khattab and Zaharia 2020), and multilingual general-purpose encoders like BGE-M3 (Xiao et al. 2024). Multi-stage refinement adds cross-encoder rerankers (Nogueira and Cho 2019) and LLM-based listwise ranking approaches such as RankGPT (Sun et al. 2023). On the query formulation side, methods including HyDE (Gao et al. 2023) expand queries via hypothetical passages. However, these methods optimize how queries are encoded, matched, and ranked against a fixed document representation. On Corpus-Side, several lines of work have modified the corpus before indexing. Early document preprocessing techniques (Manning 2008) adjust surface token forms. Early structured document retrieval (Wilkinson 1994) exploited pre-existing field markup for weighted term scoring but relied on author-provided structure. Extractive summarization for IR (Tombros and Sanderson 1998) selects a subset of existing sentences and trades recall for conciseness without generating new, denser text. Document expansion via query prediction such as Doc2Query (Nogueira et al. 2019), appends synthetic queries to broaden lexical coverage for term matching, yet leaves the document’s original knowledge content and information density unchanged. Chunking strategies (Bhat et al. 2025) determine where to place segment boundaries but preserve the token composition within each segment. Across these approaches, a common pattern emerges: each modifies how text is packaged, not what the text conveys. Index-Side Restructuring: Recent Progress and Remaining Gap. Recent methods have begun exploiting document structure for indexing. Multi-view Content-aware Indexing (Dong et al. 2024), DISRetrieval (Chen et al. 2026b), and PaperRegister (Wang et al. 2026) organize documents by content structure, discourse, or register but derive representations from original tokens without transformation. Summary-Augmented Chunking (Reuter et al. 2025) injects document-level context into chunks, improving retrieval without rewriting the indexed text. RAPTOR (Sarthi et al. 2024) generates recursive LLM summaries but relies on embedding-driven clustering. GraphRAG (Edge et al. 2024) and follow-up graph-based methods such as PathRAG (Chen et al. 2026a) and LeanRAG (Zhang et al. 2026) reorganize at the graph level without modifying indexed text. Across these methods, structure is exploited but text is not transformed; text is enriched but its information density is not increased. Singh et al. (2026) add a critical constraint: once a strong cross-encoder reranker is present, retrieval enhancements yield no Holm-corrected gain. PRIME occupies the resulting gap; generative text restructuring that increases information density (SR), combined with deterministic metadata extraction (MS) and structure-aware hierarchical retrieval (HR), with characterized boundary conditions. 3 Method 3.1 Preliminary Let =D1,…,DNC=\D_1,…,D_N\ denote the corpus where each document DiD_i consists of second-level headings (H2). PRIME processes documents through two tracks, illustrated in Figure 2. Semantic Restructuring (SR) employs LLM-driven analysis to transform DiD_i into a knowledge-dense representation Di∗D_i^* that preserves the original H2 structure. Metadata Synthesis (MS) and Hierarchical Retrieval (HR) form a deterministic track where MS extracts top-5 discriminative terms t and a heading tree h from the corpus (DiD_i or Di∗D_i^*), and HR consumes these metadata through a two-stage pipeline: coarse document filtering on high-density metadata embeddings followed by section-level cross-encoder reranking. PRIME supports two deployment modes. Offline mode runs MS and HR directly on the original corpus without SR, producing metadata Mi=MS(Di)M_i=MS(D_i). Full mode applies SR to produce Di∗D_i^*, then extracts metadata from the restructured text (Mi∗=MS(Di∗)M_i^*=MS(D_i^*)) and runs HR; Di∗D_i^* also replaces DiD_i as input to BM25 and the cross-encoder reranker. For component ablation, we also evaluate the SR body text by itself without MS metadata (denoted SR (full-text) in Table 2). Formally, the two modes are: Offline: : Di→MSMi D_i MSM_i (1) Ivec=Index(Mi) I_vec=Index(M_i) Ibm25=Index(Di) I_bm25=Index(D_i) Full: : Di→SRDi∗ D_i SRD_i^* Di∗→MSMi∗ D_i^* MSM_i^* Ivec=Index(Mi∗) I_vec=Index(M_i^*) Ibm25=Index(Di∗) I_bm25=Index(D_i^*) where Mi=(ti,hi)M_i=(t_i,h_i) denotes the top-5 discriminative terms and heading tree. 3.2 Semantic Document Restructuring SR transforms each document DiD_i into a knowledge-dense representation Di∗D_i^* while preserving the original H2 structure. It operates along two semantic axes: Explication reconstructs implicit argumentative structures within each H2 section into declarative statements, and Compression eliminates descriptive prose while retaining core contributions, conclusions, key terms, and numerical results. By removing low-IDF filler tokens, SR shifts Di∗D_i^* toward high-IDF discriminative terms, so the cross-encoder’s pairwise attention operates on a denser semantic signal per token and the LLM generator consumes fewer tokens for the same answer-relevant content. To ensure complete coverage, SR invokes the LLM per H2 section following the document’s heading tree, mitigating lost-in-the-middle effects (Liu et al. 2024). The computational cost is c(D)=∑h∈H2(D)(α⋅len(h)+β)c(D)= _h∈ H2(D)(α·len(h)+β), where α captures per-token inference cost and β the per-section prompt and API overhead; output tokens add a symmetric cost factor (full break-even analysis in the Supplementary Material, Appendix D.6). 3.3 Deterministic Metadata Synthesis The Metadata Synthesis (MS) module generates two types of metadata for the HR pipeline (Hierarchical Retrieval): discriminative terms for document-level separation and a heading tree for section-level navigation. Discriminative terms are selected as the top-5 candidates using a TF-IDF scoring function (Salton and Buckley 1988): score(t)=TF(t)×IDFcorpus(t)×(1−α)first_pos/n×(1+β⋅in_heading(t)).score(t)=TF(t)×IDF_corpus(t)\\ ×(1-α)^first\_pos/n× (1+β·1_in\_heading(t) ). (2) where TF(t)TF(t) and IDFcorpus(t)IDF_corpus(t) denote the term frequency and corpus-level inverse document frequency of t, first_posfirst\_pos is the token position of t’s first occurrence in DiD_i, n=|Di|n=|D_i|, α=0.3α=0.3 controls the positional decay rate, and β=2.0β=2.0 weights the heading bonus. Terms are whitespace-delimited tokens after lowercasing and punctuation removal. IDFcorpusIDF_corpus uses the smoothed form logN+1df(t)+1+1 N+1df(t)+1+1 to avoid negative values when df(t)=Ndf(t)=N, and is computed from the evaluation corpus itself rather than from an external reference, since domain-specific IDF better captures distinctiveness across structurally diverse corpora (evaluated in Experiments). The heading tree, parsed from H2 headings of DiD_i, together with the document title and discriminative terms form a high-density metadata embedding for coarse filtering and provide section boundaries for fine-grained reranking. 3.4 Structure-Aware Hierarchical Retrieval The Hierarchical Retrieval (HR) module consumes MS-generated metadata through a two-stage pipeline: coarse document filtering followed by section-level reranking. This design reflects the observation that document-level relevance can be assessed via coarse semantic signals, while section-level relevance requires fine-grained token interaction. Stage 1 — High-Density Document Filtering Stage 1 retrieves candidate documents using two parallel pathways. The BM25 pathway indexes the full document text for exact keyword matches. The vector pathway indexes the high-density metadata (title + discriminative terms + heading tree) encoded via BGE-small-en-v1.5 (384-dim; Xiao et al., 2024). Results from both pathways are merged via Reciprocal Rank Fusion (RRF, k=60k=60) (Cormack et al. 2009) to form a top-30 candidate set. The heading tree is invariant to whether SR is applied (H2 structure is preserved). In PRIME (offline), discriminative terms are extracted from DiD_i. In full PRIME, terms are extracted from Di∗D_i^*, reflecting SR’s vocabulary shift; the resulting impact is quantified in the Ablation. Stage 2 — Section-Level Evidence Localization Stage 2 scores each H2 section of candidate documents using a cross-encoder (ms-marco-MiniLM-L-6-v2; Reimers & Gurevych, 2019), fine-tuned on MS MARCO (Bajaj et al. 2016). The top-5 sections by CE score are forwarded to the LLM generator. Document-level rankings in Experiments are derived via best-section aggregation: score(D)=maxs∈DCE(q,s)score(D)= _s∈ DCE(q,s). The lightweight cross-encoder eliminates query-time LLM overhead typical of listwise rerankers. Full model configuration and hyperparameters are provided in Appendix B. 4 Experiments Datasets We evaluate on three structurally diverse benchmarks: QASPER (Dasigi et al. 2021) serves as the target scenario for PRIME’s full pipeline: we use 416 papers from the official test set, producing 1,451 queries with document-level qrels (one relevant document per query, 415 negatives). Queries are natural language questions without title or metadata. BRIGHT (SU et al. 2025) tests cross-domain generalization of PRIME’s offline mode across mixed-structure sources. We evaluate all twelve subsets despite four (AoPS, LeetCode, TheoremQA-Q, and TheoremQA-Theorems) requiring reasoning which fall outside PRIME’s design scope. NFCorpus (Boteva et al. 2016), from the BEIR benchmark (Thakur et al. 2021), defines the boundary condition where structural metadata provides no discriminative signal. Evaluation Metrics We evaluate PRIME across four complementary dimensions: retrieval quality, downstream QA utility, text density, and operational cost. This evaluation framework establishes a complete evidence chain from index-side representation quality to end-to-end task performance. Retrieval Quality. All three datasets use NDCG@5, P@5, R@5, P@1, R@1, and MRR. QASPER uses document-level qrels (one relevant doc per query), so P@1 ≡ R@1 and R@5 = 5×P@5, and NDCG@5 reduces to a rank-based metric driven entirely by the position of the single relevant document. NFCorpus uses graded relevance (0–2); BRIGHT uses evidence-based binary qrels. Consumption-Side QA Quality. QA is evaluated via QASPER’s official Answer-F1 (Dasigi et al. 2021), computed as the best match over the three reference answers per question. Text Density — Corpus-Level Transformation. We quantify SR’s effect on document representation via four metrics. Three capture orthogonal dimensions: stopword ratio (filler compression), technical term density (domain terms per 100 tokens), and mean IDF (average token distinctiveness) (Robertson 2004). The fourth, semantic density, integrates them into a composite signal-to-noise ratio. We define semantic density as: Semantic Density=content_tokens×mean_idfstop_tokens+1Semantic Density= content\_tokens×mean\_idfstop\_tokens+1 The numerator estimates total information carried by a document, content-bearing tokens weighted by their average distinctiveness across the corpus. The denominator estimates noise, filler words that contribute surface length without semantic content. A full derivation appears in Appendix D.7. Implementation All PRIME variants share a unified backend: BGE-small-en-v1.5 (384-dim) (Xiao et al. 2024), hnswlib configured with M=32, ef=200, and cosine similarity (Malkov and Yashunin 2018), and SQLite FTS5 with porter stemming. All query-side components (embedding, BM25, vector search, and cross-encoder (ms-marco-MiniLM-L-6-v2) (Reimers and Gurevych 2019)) run on a single NVIDIA RTX 5090 GPU. SR uses DeepSeek-V4-pro (Xu et al. 2026) at temp=0; QA generation in Context Quality uses DeepSeek-V4-Flash. All hyperparameters were fixed before evaluation based on literature defaults with no tuning on test queries. Full configuration and hyperparameter tables are in Appendix D. 4.1 Main Results Retrieval Quality QASPER Index-side optimization improves retrieval, but its measurable advantage narrows as downstream pipeline components increasingly compensate for weaker embeddings. We establish this through two complementary analyses: a shared-pipeline ceiling that reveals convergence under a strong reranker (Table 1), and a factorial decomposition that isolates index representation from pipeline effects (Table 2). Table 1: Retrieval quality on QASPER 416 (1,451 queries). BGE-M3 uses chunk-256 max-pooling; Summary-Augmented prepends section summaries to chunks; RAPTOR-lite does section+window retrieval without LLM summarization. Structure-chunks is a simpler H2-aware chunking baseline (no MS metadata, no HR pipeline). PRIME (Full) is excluded because SR alters the BM25-indexed text, breaking the shared-pipeline condition. Method NDCG5 NDCG10 Recall5 Precision5 P1 MRR Raw (chunk-256) 0.391 0.405 0.462 0.092 0.306 0.373 BGE-M3 (1024-dim) 0.389 0.398 0.453 0.091 0.306 0.371 Summary-Augmented 0.398 0.411 0.471 0.094 0.309 0.379 RAPTOR-lite 0.391 0.408 0.461 0.092 0.307 0.375 PRIME (offline) (ours) 0.396 0.409 0.469 0.094 0.309 0.377 Structure-chunks 0.396 0.411 0.466 0.093 0.310 0.379 All six methods converge within NDCG@5 = 0.389–0.398, echoing the reranker-saturation effect reported by Singh et al. (2026): once a strong cross-encoder is in the loop, corpus-side differences become nearly inseparable. This apparent equivalence conflates index quality with pipeline strength. We therefore evaluate a 5×35× 3 factorial matrix varying only the indexed text representation across three pipelines (Dense-only, Hybrid BM25+RRF, Hybrid+CE), disentangling the two. 95% paired bootstrap CIs (10,000 resamples) are reported. Table 2: Factorial ablation: NDCG@5 on QASPER 416. 95% paired bootstrap CIs in brackets. All configurations share identical retrieval pipelines; only the indexed text varies between rows. The five index representations and three retrieval pipelines are defined in Preliminary. Index Representation Dense-only Hybrid (BM25+Vec+RRF) Hybrid+CE Raw (chunk-256) 0.175 [.164,.186] 0.311 [.296,.327] 0.391 [.370,.412] Structure-chunks 0.201 [.187,.216] 0.338 [.321,.354] 0.396 [.375,.416] PRIME (offline) 0.201 [.186,.217] 0.332 [.315,.349] 0.396 [.375,.416] SR (full-text) 0.208 [.196,.221] 0.245 [.232,.258] 0.328 [.309,.347] PRIME (Full) 0.214 [.199,.229] 0.241 [.227,.256] 0.325 [.307,.345] As shown in Table 2, Index representation matters in isolation: in the Dense-only column, all four structured representations outperform chunk-256 (p<0.001p<0.001). The convergence in table 1 reflects that BM25 and the CE reranker mask these underlying embedding differences: the advantage of structured over raw narrows from +15–22% (Dense-only) to +7–9% (Hybrid) to <0.005 (Hybrid+CE). Within this narrowing gap, structured representations exhibit distinct efficiency trade-offs: Deterministic metadata matches structure-aware chunking at a fraction of the cost. Structure-chunks and PRIME (offline) remain indistinguishable (overlapping 95% CIs). Yet, PRIME (offline) embeds only metadata while capturing the same document-filtering signal as full-text chunking, at 45.8× less embedding computation and 22.4× fewer indexed vectors (Operational Cost). In contrast, both SR-based representations trade lexical signal for semantic density. PRIME (Full) achieves the highest Dense-only NDCG@5 (0.214), confirming that the compact metadata embedding (title, topic tags, and heading tree extracted from SR text, ~200 tokens) already captures stronger discriminative signal than raw chunk-256. However, BM25—common to both SR (full-text) and PRIME (Full) via the same SR body index—drops substantially below Raw in the Hybrid pipeline (0.241 and 0.245 vs. 0.311), because rewriting shifts text toward high-IDF terms at the expense of surface keyword overlap. CE reranking partially recovers this loss (Hybrid+CE: 0.325 and 0.328), though this retrieval-side cost is ultimately offset by improved context quality during generation (Context Quality). BRIGHT We evaluate PRIME (offline) on all twelve BRIGHT tasks. Table 3a reports the eight retrieval tasks; TheoremQA-Theorems is excluded from the average and noted in the caption. The other three reasoning tasks are reported separately in Appendix E of Supplementary Material. Table 3: Cross-domain retrieval on the eight BRIGHT structured tasks. Δ = (PRIME-Baseline)/Baseline; 95% CIs from 10,000-sample paired bootstrap (BCa) over per-query NDCG@5. Per-task CIs span zero (65-103 queries/task), so individual task deltas are not independently significant; the +6.9% macro-average reflects a distributional shift across domains rather than a per-task claim. TheoremQA-Theorems (PRIME NDCG@5 = 0.012, Δ = -59.8%) is excluded from this average. Tasks BGE-small (baseline) PRIME (offline) Δ NDCG@5 95% CI (Δ ) NDCG@5 P@5 R@5 NDCG@5 P@5 R@5 Pony 0.067 0.073 0.048 0.185 0.154 0.126 +175.6%+175.6\% [−0.063,+0.299][-0.063,\ +0.299] Biology 0.379 0.128 0.491 0.464 0.155 0.598 +22.3%+22.3\% [−0.188,+0.357][-0.188,\ +0.357] Stackoverflow 0.234 0.075 0.359 0.272 0.079 0.363 +16.2%+16.2\% [−0.203,+0.279][-0.203,\ +0.279] Robotics 0.181 0.052 0.253 0.200 0.059 0.282 +10.5%+10.5\% [−0.250,+0.290][-0.250,\ +0.290] Economics 0.247 0.070 0.333 0.258 0.074 0.354 +4.6%+4.6\% [−0.242,+0.265][-0.242,\ +0.265] Sustainable_living 0.250 0.080 0.342 0.242 0.082 0.343 −3.1%-3.1\% [−0.259,+0.244][-0.259,\ +0.244] Psychology 0.217 0.069 0.304 0.199 0.061 0.259 −8.2%-8.2\% [−0.277,+0.242][-0.277,\ +0.242] Earth_science 0.458 0.157 0.537 0.352 0.128 0.419 −23.1%-23.1\% [−0.404,+0.193][-0.404,\ +0.193] Avg (8 tasks) 0.254 0.088 0.333 0.272 0.099 0.343 +6.9%+6.9\% — The Δ distribution across the eight structured tasks falls into three qualitative regimes, each predicted by corpus properties that can be assessed before retrieval. At one extreme, deficit-bridging (pony, +175.6%): dense API reference documentation is poorly represented by BGE-small’s pretraining distribution, so the discriminative terms by MS serve as a natural-language bridge for the encoder. Typical structured documents earn modest gains from structure-exploiting (biology to economics, +4.6% to +22.3%), where structured H2 sections with [edit] wiki markers anchor CE section reranking. At the opposite end, structure-absent or saturated domains (sustainable_living to earth_science, -3.1% to -23.1%) show that prose already well-embedded by BGE-small, with sparse or empty heading trees, receives no additional discriminative signal from metadata. TheoremQA-Theorems (-59.8%, excluded from the average) marks a distinct fourth mode rather than saturation: its LaTeX blocks are parsed as H2 sections but fragment mathematical proofs and requires reasoning, so the CE scores an isolated definition against a query that requires a complete inferential chain. NFCorpus NFCorpus (Boteva et al. 2016) defines the opposite boundary. This dataset features short documents without H2 headings, PRIME extracts only a pseudo-title with five TF-IDF topic tags per document. We also report PRIME (Full-Text) where the same topic tags are injected alongside the original full-text embedding to isolate the metadata contribution. Table 4: Retrieval quality on the full NFCorpus. Graded relevance 0–2 binarized at ≥1≥ 1. PRIME(FT) denotes PRIME (Full-Text), Δ1 _1 denotes PRIME vs BGE-small and Δ2 _2 denotes PRIME (Full-Text) vs BGE-small. Method NDCG@5 P@5 R@5 MRR BM25 0.350 0.295 0.119 0.530 BGE-small 0.374 0.324 0.121 0.538 BGE-base 0.387 0.336 0.126 0.561 PRIME 0.301 0.264 0.084 0.460 PRIME (FT) 0.366 0.316 0.118 0.540 Δ1 _1 -19.5% -18.5% -30.6% -14.5% Δ2 _2 -2.1% -2.5% -2.5% +0.4% As shown in Table 4, PRIME alone drops 19.5% NDCG@5 below BGE-small: the five topic tags are insufficient context for short, high-density abstracts. In contrast, PRIME (Full-Text) recovers to within 2.1% of BGE-small (NDCG@5 0.366 vs 0.374), slightly exceeding in MRR, confirming that the tags introduce a mild perturbation without adding discriminative signal, the boundary condition where structural metadata ceases to help and begins to interfere. Context Quality Retrieval Quality showed that SR degrades retrieval under Hybrid+CE (NDCG@5 0.325 vs. 0.391 for Raw). We find that SR’s downstream gains in context quality partially compensate for this retrieval cost. Table 5 traces the causal chain from text density through evidence localization to answer quality. Table 5 quantifies SR’s effect at two intermediate stages. Panel A reveals that SR increases the knowledge density of the context substantially: filler tokens drop by 36.3%, mean IDF rises 19.7%, and semantic density increased by 127.3%. Panel B shows that this signal propagates to all evidence retrieval metrics on the 708 queries with section-level evidence in both corpora: SR+CE outperforms Raw+CE on section-level NDCG@5 (+45.7%), Recall@5 (+59.9%), Precision@5 (+7.9%), MAP (+69.4%), MRR (+15.4%), and Token Recall@5 (+14.1%), while indexing 29.0% fewer evidence tokens per query, confirming that the CE better localizes relevant content within fewer tokens when the underlying text is denser. Across QASPER’s question categories, SR improves section-level Recall@5 for extractive (+63%), abstractive (+56%), and unanswerable (+61%) queries, indicating the gain stems from text density rather than answer format. Yes/no queries are excluded (2 queries with evidence-bearing sections). Full derivation of density metrics in Appendix D.5, full evidence retrieval metrics in Appendix D.7. Table 5: SR causal chain: text density and evidence localization (QASPER 416). Panel A uses fixed raw-corpus reference IDF. Panel B uses section qrels from gold evidence spans mapped to H2 sections, averaged over the 708 shared queries of Raw and SR section qrels to avoid survivor bias. Raw SR Δ % Panel A: Corpus-level density (416 docs) Stopword ratio 0.365 0.233 -36.3% Mean IDF 1.947 2.331 +19.7% Tech term density (per 100 toks) 46.3 53.8 +16.1% Semantic density (composite) 3.42 7.76 +127.3% Panel B: Section-level evidence retrieval (CE reranker) NDCG@5 0.127 0.185 +45.7% Recall@5 0.128 0.204 +59.9% Precision@5 0.089 0.096 +7.9% MAP 0.095 0.161 +69.4% MRR 0.198 0.229 +15.4% Token Recall@5 0.442 0.505 +14.1% Gold evidence toks per query 346.9 246.3 -29.0% Finally, we quantify SR’s end-to-end QA impact, denoting configurations as retrieval corpus to generation corpus. All configurations share the same full-body hybrid retrieval pipeline, differing only in whether the raw or SR corpus being indexed and read. Raw to Raw yields a baseline Answer-F1 of 0.0842. Raw to SR adds +2.9% (0.0866), isolating SR’s value as denser reading material. Full to Full reaches 0.0957 (+13.7%), despite its document-level retrieval being substantially weaker (demonstrated in Table 2). The gap between Raw to SR and Full to Full shows that SR benefits QA primarily through improved cross-encoder evidence localization on denser sections, which more than compensates for the document-level retrieval drop. SR Fidelity SR increases text density but risks discarding answer-critical content or introducing factual errors; Table 6 summarizes three audit dimensions across evidence retention, entity preservation, and NLI entailment on QASPER (416 documents, 5,589 raw to SR section pairs, 2,231 gold evidence spans). Table 6: SR fidelity audit on QASPER. Panel A: a span is retained if its character-level content appears verbatim in the SR block. Panel B uses spaCy en_core_web_sm. Panel C uses roberta-large-mnli. Metrics Value Panel A: Evidence retention (2,231 gold spans, 373 docs) Overall span retention 63.2% Median per-document retention 66.7% (p25: 44.4%, p75: 100%) Perfect retention (100% of spans) 26.8% of docs High retention (80–99%) 35.1% of docs Low retention (<50%) 25.7% of docs Panel B: Entity & signal preservation (raw→ , paired) Numerical values preserved 91.0% Acronyms preserved 92.5% Named entities preserved 72.7% Extraneous spans (hallucinated) 1.3% Panel C: NLI entailment (roberta-large-mnli, 5,589 pairs) Entailment 4,061 (72.7%) Neutral 1,286 (23.0%) Contradiction 242 (4.3%) As shown in Table 6, 63.2% of evidence spans survive restructuring; losses are almost exclusively through compression rather than hallucination (only 1.3% of spans appear in SR but not raw text). The 25.7% of documents with retention below 50% share a recognizable profile: multi-step procedural reasoning or cross-section comparisons whose intermediate steps SR’s declarative reformulation flattens into short summaries. Panel B shows strong preservation of atomic signals (numerical values 91.0% and acronyms 92.5%); named entity rates drop to 72.7%, reflecting repeated mentions collapsing under compression. Panel C reveals a 4.3% contradiction rate (95.7% NLI-consistent); 181 of 416 documents have at least one contradiction, concentrated in formal/mathematical content where declarative reformulation flattens conditional or quantified statements. Taken together, these three audits define the quality-fidelity profile of SR. For roughly 60% of documents, SR delivers substantially denser text at negligible factual cost. For the remaining documents, the primary risk is omission under compression, not fabrication. Therefore, SR is best suited to declarative, evidence-dense documents while domains rich in procedural reasoning or formal logic require caution. Operational Cost PRIME (offline) consumes zero query-time LLM tokens and matches structure-aware chunking in retrieval quality (Retrieval Quality, overlapping 95% CIs). SR adds a one-time preprocessing cost of 5̃6K tokens/doc including per-section prompt overhead (Semantic Restructuring). Against LLM pointwise reranking (~1,926 tokens/query per 20-candidate batch), SR breaks even at 12,096 queries, roughly 8.3× QASPER’s test set, after which it too operates at zero per-query LLM cost. Full resource accounting tables, per-dataset cost breakdowns, and break-even analysis are in Appendix D.6. 4.2 Ablation The 5×35× 3 factorial design of Table 2 constitutes a component-wise ablation, isolating index representations under fixed retrieval pipelines. Under Dense-only, PRIME (Full) leads at 0.214 (+22.2% over raw, p<0.001p<0.001); PRIME (offline) matches structure-chunks (0.201 vs. 0.201). Under Hybrid+CE, SR-based variants drop below raw (0.325–0.328 vs. 0.391, p<0.001p<0.001), confirming that CE reranking on unmodified H2 sections saturates document-level signal—index-side gains primarily manifest at the context level (Context Quality). 5 Discussion Our results reveal that index-side optimization is effective at the context level. However, the fact that deterministic metadata matches structured chunking while SR’s QA gains flow primarily through improved retrieval rather than context substitution points to a broader issue: retrieval metrics alone are an incomplete lens on corpus-side preprocessing. Additionally, the three-regime pattern across BRIGHT points to a broader implication. Index-side interventions are not universally beneficial, their utility is predictable from corpus properties measurable before retrieval. This shifts the evaluation focus from absolute improvement to corpus-conditioned utility, precisely the dimension that standard IR benchmarks, averaged over heterogeneous corpora, cannot resolve. Four open problems follow. SR’s quality-fidelity trade-off raises the question of when rewriting can be trusted without verification; its break-even cost (Operational Cost) motivates lighter restructuring strategies without losing information; the interference on heading-free corpora (Retrieval Quality, NFCorpus) suggests corpus-adaptive strategies; and our text-only pipeline is inherently limited to textual evidence — table and figure content in QASPER is not indexed. These call for a principled theory of index-side optimization conditioned on document structure, rather than empirical heuristics. 6 Conclusion We presented PRIME, an index-side optimization framework combining LLM-driven Semantic Restructuring with deterministic Metadata Synthesis and Hierarchical Retrieval. Deterministic metadata matches structure-aware chunking at 45.8× less embedding computation and 22.4× fewer indexed vectors. On top of these gains, Semantic Restructuring further improves Answer-F1 by 13.7% in full deployment, with characterized quality–fidelity–cost trade-offs. Through extensive evaluation, we identified the operating conditions and cost profiles of this framework. Future work includes improving SR fidelity under text restructuring in order to fulfill its retrieval potential. References Bajaj et al. (2016) P. Bajaj, D. Campos, N. Craswell, L. Deng, J. Gao, X. Liu, R. Majumder, A. McNamara, B. Mitra, T. Nguyen, et al. MS marco: a human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268. Cited by: §3.4. Bhat et al. (2025) S. R. Bhat, M. Rudat, J. Spiekermann, and N. Flores-Herr Rethinking chunk size for long-document retrieval: a multi-dataset analysis. arXiv preprint arXiv:2505.21700. Cited by: §2. Borgeaud et al. (2022) S. Borgeaud, A. Mensch, J. Hoffmann, T. Cai, E. Rutherford, K. Millican, G. B. Van Den Driessche, J. Lespiau, B. Damoc, A. Clark, et al. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, p. 2206–2240. Cited by: §1. Boteva et al. (2016) V. Boteva, D. Gholipour, A. Sokolov, and S. Riezler A full-text learning to rank dataset for medical information retrieval. In European Conference on Information Retrieval, p. 716–722. Cited by: §4, §4.1. Chen et al. (2026a) B. Chen, Z. Guo, Z. Yang, Y. Chen, J. Chen, Z. Liu, C. Shi, and C. Yang PathRAG: pruning graph-based retrieval augmented generation with relational paths. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 30183–30191. Cited by: §2. Chen et al. (2026b) H. Chen, Y. Yang, Y. Li, M. Zhang, B. Hu, and M. Zhang Beyond chunking: discourse-aware hierarchical retrieval for long document question answering. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, p. 18176–18198. External Links: Link, Document, ISBN 979-8-89176-390-6 Cited by: §2. Cormack et al. (2009) G. V. Cormack, C. L. Clarke, and S. Buettcher Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, p. 758–759. Cited by: §3.4. Dasigi et al. (2021) P. Dasigi, K. Lo, I. Beltagy, A. Cohan, N. A. Smith, and M. Gardner A dataset of information-seeking questions and answers anchored in research papers. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, p. 4599–4610. Cited by: §4, §4. Dong et al. (2024) K. Dong, D. G. X. Deik, Y. Q. Lee, H. Zhang, X. Li, C. Zhang, and Y. Liu Multi-view content-aware indexing for long document retrieval. arXiv preprint arXiv:2404.15103. Cited by: §2. Edge et al. (2024) D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, D. Metropolitansky, R. O. Ness, and J. Larson From local to global: a graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130. Cited by: §2. Formal et al. (2021) T. Formal, B. Piwowarski, and S. Clinchant Splade: sparse lexical and expansion model for first stage ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 2288–2292. Cited by: §2. Gao et al. (2023) L. Gao, X. Ma, J. Lin, and J. Callan Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 1762–1777. Cited by: §2. Guu et al. (2020) K. Guu, K. Lee, Z. Tung, P. Pasupat, and M. Chang Retrieval augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, H. D. I and A. Singh (Eds.), Proceedings of Machine Learning Research, Vol. 119, p. 3929–3938. External Links: Link Cited by: §1. Izacard et al. (2021) G. Izacard, M. Caron, L. Hosseini, S. Riedel, P. Bojanowski, A. Joulin, and E. Grave Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118. Cited by: §2. Izacard et al. (2023) G. Izacard, P. Lewis, M. Lomeli, L. Hosseini, F. Petroni, T. Schick, J. Dwivedi-Yu, A. Joulin, S. Riedel, and E. Grave Atlas: few-shot learning with retrieval augmented language models. Journal of Machine Learning Research 24 (251), p. 1–43. Cited by: §1. Karpukhin et al. (2020) V. Karpukhin, B. Oguz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. Yih Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), p. 6769–6781. Cited by: §1, §2. Khattab and Zaharia (2020) O. Khattab and M. Zaharia Colbert: efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, p. 39–48. Cited by: §1, §2. Lewis et al. (2020) P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, p. 9459–9474. Cited by: §1. Liu et al. (2024) N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang Lost in the middle: how language models use long contexts. Transactions of the Association for Computational Linguistics 12, p. 157–173. External Links: Link, Document Cited by: §3.2. Malkov and Yashunin (2018) Y. A. Malkov and D. A. Yashunin Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence 42 (4), p. 824–836. Cited by: §4. Manning (2008) C. D. Manning Introduction to information retrieval. Syngress Publishing,. Cited by: §2. Nogueira and Cho (2019) R. Nogueira and K. Cho Passage re-ranking with bert. arXiv preprint arXiv:1901.04085. Cited by: §2. Nogueira et al. (2019) R. Nogueira, W. Yang, J. Lin, and K. Cho Document expansion by query prediction. arXiv preprint arXiv:1904.08375. Cited by: §2. Reimers and Gurevych (2019) N. Reimers and I. Gurevych Sentence-bert: sentence embeddings using siamese bert-networks. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), p. 3982–3992. Cited by: §4. Reuter et al. (2025) M. Reuter, T. Lingenberg, R. Liepina, F. Lagioia, M. Lippi, G. Sartor, A. Passerini, and B. Sayin Towards reliable retrieval in RAG systems for large legal datasets. In Proceedings of the Natural Legal Language Processing Workshop 2025, N. Aletras, I. Chalkidis, L. Barrett, C. Goanță, D. Preoțiuc-Pietro, and G. Spanakis (Eds.), Suzhou, China, p. 17–30. External Links: Link, Document, ISBN 979-8-89176-338-8 Cited by: §2. Robertson and Zaragoza (2009) S. Robertson and H. Zaragoza The probabilistic relevance framework: bm25 and beyond. Found. Trends Inf. Retr. 3 (4), p. 333–389. External Links: ISSN 1554-0669, Link, Document Cited by: §1, §2. Robertson (2004) S. Robertson Understanding inverse document frequency: on theoretical arguments for idf. Journal of documentation 60, p. 503–520. Cited by: §4. Salton and Buckley (1988) G. Salton and C. Buckley Term-weighting approaches in automatic text retrieval. Information processing & management 24 (5), p. 513–523. Cited by: §3.3. Sarthi et al. (2024) P. Sarthi, S. Abdullah, A. Tuli, S. Khanna, A. Goldie, and C. Manning Raptor: recursive abstractive processing for tree-organized retrieval. In International Conference on Learning Representations, Vol. 2024, p. 32628–32649. Cited by: §2. Singh et al. (2026) S. Singh, A. Reddy, and M. Chopra Beyond the reranker: do rag retrieval enhancements help once a strong reranker is present?. arXiv preprint arXiv:2606.28367. Cited by: §1, §2, §4.1. SU et al. (2025) H. SU, H. Yen, M. Xia, W. Shi, N. Muennighoff, H. Wang, L. Haisu, Q. Shi, Z. S. Siegel, M. Tang, R. Sun, J. Yoon, S. O. Arik, D. Chen, and T. Yu BRIGHT: a realistic and challenging benchmark for reasoning-intensive retrieval. In The Thirteenth International Conference on Learning Representations, External Links: Link Cited by: §4. Sun et al. (2023) W. Sun, L. Yan, X. Ma, S. Wang, P. Ren, Z. Chen, D. Yin, and Z. Ren Is chatgpt good at search? investigating large language models as re-ranking agents. In Proceedings of the 2023 conference on empirical methods in natural language processing, p. 14918–14937. Cited by: §2. Thakur et al. (2021) N. Thakur, N. Reimers, A. Rücklé, A. Srivastava, and I. Gurevych BEIR: a heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), External Links: Link Cited by: §4. Tombros and Sanderson (1998) A. Tombros and M. Sanderson Advantages of query biased summaries in information retrieval. In Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’98, New York, NY, USA, p. 2–10. External Links: ISBN 1581130155, Link, Document Cited by: §2. Wang et al. (2026) S. Wang, H. Zhang, J. Li, and Y. Liu PaperRegister: hierarchical register indexing for academic document retrieval. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics, Cited by: §2. Wilkinson (1994) R. Wilkinson Effective retrieval of structured documents. In SIGIR’94: Proceedings of the Seventeenth Annual International ACM-SIGIR Conference on Research and Development in Information Retrieval, organised by Dublin City University, p. 311–317. Cited by: §2. Xiao et al. (2024) S. Xiao, Z. Liu, P. Zhang, N. Muennighoff, D. Lian, and J. Nie C-pack: packed resources for general chinese embeddings. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, p. 641–649. Cited by: §2, §4. Xiong et al. (2021) L. Xiong, C. Xiong, Y. Li, K.-F. Tang, J. Liu, P. Bennett, J. Ahmed, and A. Overwijk Approximate nearest neighbor negative contrastive learning for dense text retrieval. In Proceedings of the 9th International Conference on Learning Representations (ICLR), Cited by: §2. Xu et al. (2026) A. Xu, B. Lin, B. Xue, B. Wang, B. Xu, B. Wu, B. Zhang, C. Lin, C. Dong, C. Ling, et al. Deepseek-v4: towards highly efficient million-token context intelligence. arXiv preprint arXiv:2606.19348. Cited by: §4. Zhang et al. (2026) Y. Zhang, R. Wu, P. Cai, X. Wang, G. Yan, S. Mao, D. Wang, and B. Shi Leanrag: knowledge-graph-based generation with semantic aggregation and hierarchical retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 34862–34869. Cited by: §2.