A preprint describes a way to search whole multi-page documents that scored higher than a page-by-page system while requiring a much smaller search index. DocPC-ColQwen recorded an average NDCG@5 of 44.09%, compared with 38.91% for the strongest matched page-level baseline. NDCG@5 is the study's top-five ranking score, summarizing how well relevant documents are placed near the front of the results. The reported index used 3,850 images rather than 38,746, 2.91 million vectors rather than 29.25 million, and 2.2 gigabytes of storage rather than 22.4 gigabytes. Total indexing time was 2,034 seconds, versus 15,588 seconds for the page-level system, a reported 7.7-fold speedup.
One grid replaces a page-by-page index
The task is document-level visual retrieval: a system ranks whole multi-page documents for a query by document-level relevance. DocPC selects representative pages, arranges them in a default 2 by 2 grid, and encodes that grid once for each document. The matched page-level model shares the same visual-language model and late-interaction scoring, but indexes pages independently and combines their results with MaxP, so the strongest page match sets the document score. DocPC instead scores one grid representation directly. The comparison therefore keeps the model family aligned while changing how pages are presented to the retriever.
The test was built around whole documents
Tests used DocViRe, a benchmark that groups documents by topical similarity and uses an LLM to generate document-level query-ground-truth pairs. It covers seven English domains. The training split contains 10,884 queries over 4,438 documents, while the test split contains 2,237 queries over 3,850 documents; the two document pools are disjoint. This gives the benchmark separate document pools for learning and evaluation.
The test annotations were checked by three human annotators. Reported relevance agreement was 92.5% when all three agreed and 98.8% when at least two agreed. Agreement on query validity was 99.0% across all three annotators. The paper reports these as checks of the test-set annotations and does not report an independent external annotation study.
The score gap held across domains
One comparison kept the page budget the same. In that setting, grid composition scored 6.41 NDCG@5 percentage points above independent-page MaxP, with a positive advantage in all seven domains. The largest reported gaps were 21.08 points in Industrial and 12.00 in Education. The comparison is specific to the benchmark and configuration, and no uncertainty intervals were reported for the ablation.
Training combined two objectives. DocPC fine-tuned pretrained ColQwen with multi-positive InfoNCE, which allows more than one relevant match to contribute during training, and an ApproxNDCG listwise loss, which focuses on the ordering of a result list. The multi-positive term was applied every step, while the listwise term was applied intermittently. In the objective ablation, the combined schedule with the listwise term every 40 steps reached 44.09% average NDCG@5. Multi-positive-only training scored 42.68%, listwise-only training scored 37.32%, and combined schedules with intervals of one and 50 steps scored 40.53% and 42.83%.
Compactness comes with limits
The compact representation has a detail cost. The authors say grid compression can make small text and dense tables illegible. They also explored only query-agnostic page selection, so the tested page choices were not tailored to each query.
The benchmark's reach is limited as well. DocViRe covers seven English domains, but not broader language or genre coverage or downstream retrieval-augmented generation, where retrieved material would be passed to a generative model. Whether the reported ranking gains transfer to those settings remains open.
The work is an arXiv version 1 preprint dated 26 August 2026. No separate funding source is reported, and the ethics statement says no external crowd workers were recruited and no additional compensation was involved.
Paper data and sources
Original title: DocPC: Document-Level Visual Retrieval via Representative Page Composition
Authors: Chengsong You, Junwei Zhou, Nan Du
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text