This work proposes DocPC, a document-level visual retrieval framework based on Representative Page Composition, and introduces DocViRe, a benchmark with multi-positive relevance annotations that combines multi-positive contrastive learning with sparsely scheduled listwise optimization.
Abstract
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.
Document Visual Question Answering (DocVQA) often leverages Retrieval-Augmented Generation (RAG), where late-interaction encoders are commonly used to identify document pages relevant to a user query, before answer generation by a Large Vision-Language Model (LVLM). Existing approaches typically retrieve a fixed top-$k$ number of pages regardless of query complexity, which increases LVLM latency and may degrade answer accuracy. We introduce ViSAR (Visual Semantic Activation Retrieval), a training-free adaptive-$k$ retrieval method for late-interaction visual document retrieval. ViSAR operates directly in the embedding space to construct a query-conditioned page-level similarity matrix that highlights query-relevant semantics and dynamically determines the number of pages to retrieve. Across multiple encoders and LVLMs, ViSAR retrieves compact, query-adapted page sets that reduce RAG latency by up to 58.7\%, while maintaining or improving answer accuracy compared with fixed top-$k$ and adaptive retrieval heuristics. Furthermore, we show that the similarity matrix structure correlates with answer accuracy, suggesting future directions for retrieval quality-aware document understanding.
Adrien Mialland, Marc Plantevit, Julien Gallois et al.· 0 citations
Retrieval over visually rich documents has a representation problem: important content often lives in tables, charts, figures, and layout relations that plain OCR linearizes, corrupts, or omits. ColPali-family visual retrievers address this with patch-level multi-vector indexes and late-interaction scoring, keeping image-derived retrieval on the query-time serving path. We introduce MIDR (Multimodal Indexing for Document Retrieval), a training-free framework for enrichment-augmented indexing that shifts multimodal reasoning to index time. During ingestion, a multimodal LLM converts rendered pages into verified textual fields that are indexed with BM25F and optionally fused with dense retrieval, enabling text-centric serving over multimodally grounded evidence. On ViDoRe V3, MIDR Hybrid achieves 0.6219 average nDCG across five English domains, a 23.0% relative gain over BM25, remaining competitive with ColQwen2.5. On two French-document domains, enrichment bridges English queries and French page text, lifting BM25 from 0.1532 to 0.5448 nDCG and outperforming ColQwen2.5. Across all seven domains, MIDR leads ColQwen2.5 on four while using approximately 9x smaller index memory and approximately 2x lower query latency. These results establish index-time multimodal reasoning as a compelling accuracy-deployment alternative to serving-time visual late interaction.
Debanjan Mahata, Atharva Tendle, Daniel Preoţiuc-Pietro et al.· 0 citations
Multi-modal late-interaction retrievers achieve strong retrieval on visually rich documents by representing each page as per patch embeddings and matching at the token level. However, this approach incurs high storage costs. Existing compression methods typically fix a single compression level at indexing time, limiting flexibility. We present ColSNAP (Spatial Nested Average Pooling)1, a training method that generates a nested hierarchy of compression levels directly from a backbone's patch grid. By spatially pooling patch embeddings into pro- gressively coarser tiers and training all tiers simultaneously, a single model learns to support retrieval at multiple compression levels without architectural changes. Crucially, a single encoding pass yields every tier, enabling the accuracy-storage trade-off to be configured at indexing time to match avail- able storage budgets, rather than being fixed during training. We demonstrate that models trained using ColSNAP maintain near full-resolution retrieval performance under substantial compression and that ColSNAP transfers effectively across multiple late-interaction backbones, and achieves most of its improvements via a lightweight adaptation stage applied to a pre-trained retriever.
Scene Text Retrieval (STR) aims to search images containing a given textual query within large-scale image collections. However, existing approaches are fundamentally constrained in two ways: 1) they are evaluated on narrow benchmarks that focus primarily on natural scenes; and 2) they rely on either error-prone multi-stage recognition-then-matching pipelines or localization-assisted matching strategies. To address these limitations, we introduce MuST, the first comprehensive benchmark for multi-scene and bilingual STR tasks, covering a broad range of real-world scenarios with carefully curated Chinese and English textual queries. On top of this benchmark, we propose BPE-Ret, a novel Byte-Pair Encoding (BPE)-level retrieval framework built on a simple yet powerful principle: a word is considered present in an image if and only if all of its constituent subwords are present. Concretely, BPE-Ret decomposes textual queries into BPE subwords and directly aligns them with dense visual features within a unified embedding space, thereby eliminating the need for explicit text spotting and coarse-grained word-level matching. We further enhance fine-grained alignment through two key innovations: a weighted preference learning scheme that prioritizes challenging cases to sharpen discrimination on confusable word-image pairs, and a subword inclusive-OR matching strategy that enforces constituent subword verification to enable robust retrieval beyond word-level granularity. Extensive experiments show that BPE-Ret establishes new state-of-the-art performance on both existing public STR benchmarks and our newly proposed MuST dataset, demonstrating the effectiveness and robustness of subword-level retrieval for real-world, multilingual scene text understanding.
Tong-Kun Guan, Yu-Tong Cai, Haocheng Wang et al.· IEEE Transactions on Image P...· 0 citations
Multi-vector visual document retrievers such as ColPali and ColQwen achieve strong retrieval by storing fine-grained patch embeddings, but this produces large indexes and costly late-interaction scoring. We argue that effective compression should preserve query-relevant coverage, meaning the diverse document regions that may become the strongest MaxSim match across queries, rather than selecting patches independently by salience. This view also explains why dense rendered pages are easier to compress than natural images. We introduce MarginMerge, a compression method for frozen multi-vector retrievers. It selects coverage-aware anchors, clusters document patches, and uses a lightweight shared network to synthesize one representative per cluster. Compression is performed once during indexing, while retrieval keeps the standard MaxSim interface. Across six datasets on both ColQwen2.5 and ColPali, MarginMerge achieves the highest matched query-agnostic average at 5% and 10% vector retention. Compared with the uncompressed index using the same backbone, it preserves between 97% and 99% of average nDCG@5 while reducing stored document vectors by between 90% and 95%. At 5% retention, it also reduces ranking flips relative to geometric merging on all six ColQwen2.5 datasets by approximately 41% on average. The same model transfers to unseen datasets and retention ratios without retraining.
Ailar Mahdizadeh, Aria Salari, Sohail Rajabi et al.· 0 citations
Late-interaction visual document retrievers preserve fine-grained page evidence by storing many token embeddings per page, but the resulting storage and query-time interaction costs make large-scale deployment expensive. Pooling document tokens before indexing offers a natural remedy, yet static pooling must decide which visual evidence to preserve before the query is known. We study an alternative: a heavily compressed hot-path index generates candidates, after which query-aware token budgeting operates on the original token sets of the shortlisted pages. We formulate this stage-two selection as a budgeted MaxSim coverage problem, show that a clipped version is monotone submodular, and compare coverage-only, cluster-guided, token-wise, and marginal-gain policies. On ten ViDoRe tasks with ColModernVBERT, direct static pooling reduces macro normalized discounted cumulative gain at rank five from 0.6309 without compression to 0.4738 at a thirty-two-fold pool factor. Under the same candidate-generation regime and a pool-factor-eight-equivalent reranking budget, token top-k recovers 93.93 percent of the full-token score, while greedy marginal-gain selection recovers 98.39 percent. Held-out and leave-one-dataset-out evaluations yield positive greedy improvements over token top-k on every dataset. The latency analysis reveals two useful operating points: token top-k for interactive retrieval and the naive greedy implementation as a quality upper envelope. Together, these results show that late-interaction visual retrieval benefits from query-aware allocation rather than query-agnostic pooling alone.
PS Rishi, R. Dwivedi, V. K. Kurmi· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
