Source-linked AI summary

DocPC: Document-Level Visual Retrieval via Representative Page Composition

Chengsong You, Junwei Zhou, Nan Du

arXiv:2608.25434v1cs.IR

TL;DR

Visual document retrieval remains page-centric even though many applications require ranking complete documents, and page aggregation incurs linear indexing costs while missing distributed relevance. DocPC composes representative pages into a grid, combines multi-positive and listwise training, and introduces DocViRe; DocPC-ColQwen reaches 44.09 NDCG@5 versus 38.91 for the strongest page-level baseline while reducing storage by 10.1×.

  • Problem

    Existing visual retrieval is page-centric, while complete-document retrieval faces linear indexing costs and difficulty capturing relevance distributed across multiple pages.

  • Method

    DocPC composes representative pages into a single grid image and combines multi-positive contrastive learning with sparsely scheduled listwise optimization.

  • Results

    44.09 average NDCG@5 versus 38.91 for the strongest page-level baseline, with indexed images, vectors, and storage reduced by 10.1×.

  • Takeaways & Limitations

    DocViRe establishes document-level visual retrieval with multi-positive annotations, and document-level composition outperforms page-level aggregation under the same supervision.

  • Takeaways & Limitations

    Grid composition can reduce legibility for small text and dense tables; DocViRe covers seven English domains and excludes downstream retrieval-augmented generation.

Abstract

from arXiv · show

Visual document retrieval has advanced by encoding page screenshots with vision-language models, bypassing OCR pipelines. However, existing methods remain page-centric, misaligned with real-world scenarios requiring complete document retrieval. A naive page-then-document aggregation suffers from linear indexing cost and degraded retrieval when relevance spans multiple pages. We propose DocPC, a document-level visual retrieval framework based on Representative Page Composition: selecting representative pages and composing them into a single grid image for document-level indexing, reducing indexed images, vectors, and storage by 10.1x and end-to-end indexing time by roughly 7.7x. To handle multi-positive supervision prevalent at the document level, we combine multi-positive contrastive learning with sparsely scheduled listwise optimization. We also introduce DocViRe, a benchmark with multi-positive relevance annotations. DocPC-ColQwen achieves NDCG@5 of 44.09 on DocViRe, outperforming the strongest page-level baseline at 38.91 while reducing storage by 10.1x. Code is available at https://anonymous.4open.science/r/DocPC-Document-Level-Visual-Retrieval-via-Representative-Page-Composition-1D52. Data is available at https://huggingface.co/datasets/anonymous-7219/docpc.

1 Introduction

Existing visual document retrieval is largely page-centric, although practical use cases require complete-document retrieval. DocPC addresses this gap by composing representative pages into one document representation and adapting training and benchmarking to document-level multi-positive retrieval.

  • Document-level retrieval is underexplored despite enterprise search, academic retrieval, and legal review requiring complete documents.
  • Page-then-document aggregation scales index size linearly and can miss relevance distributed across multiple pages.
  • Representative Page Composition selects K representative pages and arranges them into one grid image, reducing document-side indexing from O(N) to O(1).
  • DocPC combines multi-positive contrastive learning with sparsely scheduled ApproxNDCG-based listwise optimization for multi-positive supervision.
  • DocViRe provides a document-level visual retrieval benchmark with multi-positive annotations.
  • DocPC is empirically validated on DocViRe, with page selection and multi-positive training producing retrieval gains while reducing index size.

2 Related Work

Visual retrieval systems commonly encode page screenshots with vision-language models and late-interaction matching, but benchmarks and methods remain focused on pages rather than whole documents. DocPC targets the underexplored cross-document setting of ranking complete multi-page documents.

  • Vision-language visual retrieval encodes page images and matches queries through late interaction over token-level embeddings.
  • Existing visual retrieval benchmarks largely evaluate page-level retrieval, intra-document page retrieval, or question answering rather than whole-document ranking.
  • Current benchmarks lack a standard setting for ranking whole documents from a corpus of multi-page documents.
  • Page-to-document aggregation methods such as MaxP reuse smaller-unit retrieval ideas, while multi-page understanding has mainly been studied for long-document question answering and reasoning.
  • The paper distinguishes cross-document retrieval, intra-document page retrieval, and multi-page document QA, targeting the first setting.

3 Task Formulation

The task ranks relevant multi-page documents for a text query, requiring compact document representations and objectives that accommodate multiple relevant documents. Page aggregation provides a baseline but does not reduce index size and can underrank documents whose relevance spans pages.

  • Document-level visual retrieval scores and ranks a corpus of multi-page documents for a text query, with relevance defined at the document level.
  • Efficient representations must avoid storage growing linearly with document page count, unlike page-level indexing with O(P_i N_i · T) storage.
  • Multi-positive supervision is necessary because several documents may be relevant to one query, making single-positive training mislabel co-relevant documents as negatives.
  • MaxP aggregates page scores into document scores but does not reduce index size and tends to underrank documents with relevance distributed across pages.
  • The first two pages have the best average relevance ranks in zero-shot ColQwen analysis, motivating early-page-aware selection strategies.

4 DocViRe Benchmark

DocViRe is a document-level visual retrieval benchmark designed around whole-document relevance, multi-positive annotations, and cross-document evaluation. Its construction combines document-level semantic grouping and query generation with LLM and exhaustive human validation.

  • Motivation: DocViRe addresses document-level retrieval, where relevance depends on topic, cross-page information, and global semantics rather than evidence from one page.
  • Benchmark design: DocViRe provides document-grounded queries and multi-positive relevance annotations for cross-document retrieval across multiple domains.
  • Statistics: The training and test sets contain 10,884 queries and 4,438 documents, and 2,237 queries and 3,850 documents, respectively, with document-disjoint splits.
  • Construction: The benchmark uses OCR and LLM keyword extraction, topical clustering, and group-wise generation of queries answered by one or more complete documents.
  • Quality validation: Human validation exhaustively reviews test samples after LLM filtering, achieving 92.5% three-way agreement for relevance and 99.0% three-way agreement for query validity.

5 Method: DocPC

DocPC converts selected document pages into a fixed-size grid representation for efficient document-level retrieval. It combines representative-page selection, late-interaction scoring, and training objectives tailored to multi-positive relevance and top-ranked results.

  • Representative Page Composition: DocPC selects K representative pages, composes them into one grid, and encodes the grid once, reducing document indexing from O(N) to O(1).
  • Page selection: First-K, Boundary, Uniform-K, Random-K, and Base-Clip provide heuristic or embedding-based strategies for selecting representative pages.
  • Grid representation: The default representation uses K = 4 pages arranged in a 2×2 grid, while K=1 and K=16 produce 1×1 and larger-grid variants.
  • Training objectives: Multi-positive InfoNCE aggregates over multiple positives so relevant documents are not treated as negatives during contrastive training.
  • Training objectives: The ApproxNDCG-based listwise loss directly focuses optimization on top-K ranking, while the combined objective introduces it intermittently alongside the multi-positive loss.
  • Inference: At inference, each document is indexed with one VLM forward pass and queried through late-interaction scores against stored multi-vector representations.

6 Experiments

Experiments evaluate DocPC on DocViRe using document- and page-level baselines, representative-page strategies, page budgets, and training objectives. DocPC improves retrieval while substantially reducing indexing resources.

  • Main results: DocPC-ColQwen (First-4) attains 44.09 average NDCG@5 versus 38.91 for the strongest page-level baseline under matched supervision.The comparison supports a document-level composition gain rather than an advantage from additional supervision.
  • Efficiency: DocPC reduces indexed images from 38,746 to 3,850, indexed vectors from 29.25M to 2.91M, indexing time from 15,588 s to 2,034 s, and storage from 22.4 GB to 2.2 GB.These correspond to approximately 10.1× reductions in indexed images, vectors, and storage, and a 7.7× reduction in end-to-end indexing time.
  • Ablation studies: Grid composition outperforms the strongest independent-page variant, MaxP, by +6.41 NDCG@5 under the same page budget.The matched comparison uses the same four selected pages and contrasts grid encoding with separate-page scoring.
  • Ablation studies: The grid advantage is positive in all seven domains, with the largest gains in Industrial (+21.08) and Education (+12.00).These gains are reported in the domain-level analysis.
  • Ablation studies: A 4-page budget is the most reasonable choice in this setting because more pages do not consistently help.The page-budget configurations use Combined (m=5) and are not directly comparable to Table 2.
  • Ablation studies: The combined objective peaks at 44.09 when the listwise loss is applied every m=40 steps.The listwise objective alone reaches 37.32, multi-positive contrastive loss alone reaches 42.68, and other schedules perform worse.

7 Conclusion

The paper proposes DocPC for efficient document-level visual retrieval and DocViRe for evaluating document-grounded retrieval with multi-positive annotations. DocPC-ColQwen improves retrieval over page-level aggregation while reducing indexing resources.

  • Contributions: DocPC composes representative pages into 2×2 grids for efficient document indexing, while DocViRe provides document-grounded queries and multi-positive annotations.The benchmark and framework are presented as the paper’s central contributions.
  • Results: DocPC-ColQwen achieves 44.09 average NDCG@5 versus 38.91 for the strongest page-level baseline on DocViRe.The result is reported under matched supervision.
  • Results: DocPC reduces indexed images, vectors, and storage by 10.1× while achieving the reported retrieval improvement.The conclusion attributes the result to document-level composition being more effective than page-level aggregation under the same supervision.

8 Limitations

DocPC’s grid representation can reduce the legibility of fine-grained page details, and DocViRe’s coverage is limited in language, genre, and downstream settings.

  • Representation limits: Grid composition reduces per-page resolution, so small text and dense tables can lose legibility.The paper suggests hybrid grid retrieval plus page reranking for detail-sensitive applications.
  • Benchmark scope: DocViRe covers seven English domains, while broader language and genre coverage remains future work.The stated domain coverage limits how broadly the benchmark’s findings can be generalized.
  • Benchmark scope: The benchmark does not yet cover downstream settings such as retrieval-augmented generation.This is an explicit scope boundary of the current benchmark.

9 Ethics Statement

The paper describes the benchmark’s data provenance and review process, and clarifies terminology and evaluation reporting conventions used in its experiments.

  • Data and review: DocViRe is constructed from the publicly available PDFA dataset of openly accessible PDF documents.Sampled documents were manually inspected without identifying personally identifiable information or offensive content.
  • Data and review: Query generation used DeepSeek V3 followed by author validation, without external crowd workers or additional compensation.The stated process combines automated generation with author review.
  • Terminology: The appendix distinguishes page-level retrieval with separate-page indexing and score aggregation from document-level retrieval with one grid encoded per document.It also distinguishes First-4 pages from First-4 grid according to whether pages are aggregated or composed.
  • Evaluation reporting: All NDCG@5 values are reported as percentages from 0 to 100.This convention applies to the reported evaluation values.

A.3 Extended Ablations on Training Objective and Page Budget

Extended ablations examine loss-frequency effects, representative-page strategies, indexing efficiency, resolution, candidate pools, and benchmark construction. Results emphasize controlled-comparison caveats alongside gains from sparse listwise refinement and document-level composition.

  • Ablation Scope: Extended loss-frequency ablations are exploratory and not directly comparable to the controlled main-text tables.The training setup and configurations may differ from the main paper.
  • Training Objective: Sparse listwise scheduling is presented as ranking refinement after multi-positive contrastive learning shapes the representation space.The listwise term is introduced intermittently, while the multi-positive loss is applied at every step.
  • Efficiency: DocPC reduces indexed images and vectors by about 10.1×, lowers Milvus storage from 22.4 GB to 2.2 GB, and cuts indexing time from 15,588 s to 2,034 s.The composed grid slightly increases per-image encoding time, but fewer indexed inputs reduce total encoding and insertion costs.
  • Resolution: Average NDCG@5 increases from 6.68 at 128×166 resolution to 40.65 at the original-resolution setting, with diminishing gains as resolution increases.Intermediate values are 22.13 at 256×331 and 37.55 at 512×662.
  • Page Selection: First-4 is the best strategy for 39.7% of individual queries, while Uniform-4 is better suited to documents with content distributed across pages.Boundary is described as effective for report-style documents with executive summaries and conclusions.

B.8 Qualitative Analysis

Qualitative analysis shows that grid composition helps when relevance spans pages but can lose fine-grained visual information. Across 2,237 queries, grid-only successes exceed page-only successes, while long documents beyond the selected pages remain challenging.

  • Outcome Distribution: 17.8% of queries are grid-only successes versus 9.7% page-only successes, yielding a 1.83× net positive ratio.A method succeeds when it retrieves at least one relevant document in the top 5.
  • Practical Implication: A hybrid approach combining grid-level retrieval with page-level reranking is proposed for applications requiring fine-grained visual detail.This follows the trade-off between cross-page context and per-page resolution.
  • Cross-Page Relevance: Grid composition is most useful for cross-page themes whose evidence co-occurs across headers, policy text, and flowcharts.The 2×2 grid can preserve these signals jointly when no single page contains sufficient evidence.
  • Fine-Grained Detail: Grid compression loses information for fine-grained visual details such as dosage tables and dense patient intake forms.Independent page encoding can preserve per-page resolution and retrieve the correct document in these cases.
  • Page Selection Limits: Queries involving documents of at least 30 pages can defeat both methods when relevant content lies beyond the first four pages.This limitation is attributed to the First-4 page-selection strategy rather than grid representation itself.
Loading 2608.25434v1…