Visual document retrieval data: slides, scans and image-heavy PDFs

Why page images need their own retrieval data

Text-only pipelines lose the evidence that sits in charts, diagrams, slide layouts and scanned pixels, so a retriever trained or tested only on extracted text will look better on paper than in production. Enterprise RAG corpora routinely mix PDF, PPTX and DOCX with HTML and Markdown [1], and a slide whose message is a bar chart often yields a title and three axis labels after parsing.

As of October 2026, many visual retrievers embed whole page images, often as multi-vector representations scored by late interaction, and skip OCR and layout parsing; making those indexes efficient is an active research problem [7]. Because OCR quality varies so much across copiers, scans and born-digital files, test both pipelines on your own pages rather than trusting one leaderboard.

The practical consequence for buyers: you need data that lets you measure where pixels beat text and where they do not. That means one corpus, two representations, and queries written to stress the visual side. For the broader retrieval buying framework, start at the retrieval and RAG data hub.

What public benchmarks cover, and where they stop

Public sets are good for method selection but rarely match an enterprise page mix, and each one isolates a slice of the problem. Check the license of each before any commercial use; our guide to public retrieval datasets and commercial-use licenses walks through that review.

Two failure modes recur. Benchmarks drift toward saturation as top models are tuned against them, so a public score says less each year about your corpus. Public test sets also carry label noise; a study of widely used benchmarks estimated an average test-set label error rate of at least 3.3% [8], so small leaderboard gaps can sit inside the error bars.

What an enterprise visual retrieval corpus should contain

The corpus you license should look like the pages your system will index, including the ugly ones. In practice that means quarterly business review decks with embedded Excel charts, engineering diagrams exported to PDF, scanned purchase orders and signed contracts, and marketing one-pagers with text baked into images.

Specify the page mix by modality, not by file extension. A PDF can be a born-digital report with a clean text layer, a scan with no text layer, or a scan with a poor OCR layer from a copier; each behaves differently under a text retriever. Ask for the share of pages that are chart-dominant, diagram-dominant, table-dominant, photo or screenshot, and text-dominant.

Ask for realistic distractors. Superseded deck versions, templated slides that differ only in one chart, and near-identical monthly reports are where page-image retrievers confuse neighbors; see distractor and near-duplicate documents for retrieval testing. Long PDFs deserve their own handling, covered in multi-page and long document data.

If your target is generating slides or training a document-understanding model rather than retrieval, the owner pages for presentation deck datasets and training data for multimodal document models are a better fit.

Writing queries that actually test visual retrieval

Queries must be answerable only, or most reliably, from visual content, otherwise a text retriever will match an image retriever and the test tells you nothing. Tag each query with the evidence type so you can slice results.

• Chart-value queries: "Which region's Q3 bar exceeded 40%?" where the value appears only as a plotted bar.
• Diagram-relation queries: "Which service sits between the API gateway and the billing queue?" answered by arrows, not prose.
• Layout queries: "The slide with the three-column pricing comparison," targeting structure rather than words.
• Scan-only queries: a term that appears in a stamped or handwritten field the OCR layer missed.
• Text-control queries: a share of queries answerable from extracted text, so the comparison has a baseline.

Write queries the way users ask, not by copying the slide title; title-copied queries inflate lexical retrievers. For multi-hop questions across slides, record every page needed, as SlideVQA does with multiple images [2]. Judging guidance, depth and assessor choice are covered in buying relevance judgments (qrels).

Delivery format: page images, text layer and page-level qrels

Deliver every page twice, as a rendered image and as extracted text, keyed to the same document ID and page number, so text, image and hybrid retrievers run on an identical corpus. A per-page JSON Lines manifest (UTF-8, one JSON object per line [9]) keeps that alignment auditable.

Fix rendering parameters in the contract: DPI (for example 144 or 200), color mode, and whether slides render with speaker notes. Record the extraction tool and version for the text layer, and keep bounding boxes for the regions that answer each query so you can later test patch-level or region-level grounding.

Illustrative example: invented to show structure; it does not describe an available dataset.

{"doc_id": "deck_0412", "page": 7, "source_format": "pptx", "image_path": "pages/deck_0412_p007.png", "render_dpi": 144, "text_layer": "native", "text_path": "text/deck_0412_p007.txt", "ocr_engine": null, "page_modality": "chart", "version_group": "qbr_2025_q3", "is_latest_version": false, "deidentification": "names_replaced"}
{"query_id": "q_1183", "query": "Which region missed its renewal target in Q3?", "evidence_type": "chart_value", "relevant": [{"doc_id": "deck_0412", "page": 7, "grade": 2, "bbox": [412, 220, 980, 610]}], "text_answerable": false}

A checklist for the delivery spec:

1. Page images at a stated DPI and format (PNG or lossless WebP), one file per page.
2. Text layer per page with its source marked native, OCR or none.
3. Page-level qrels with graded relevance and evidence bounding boxes.
4. Query tags: evidence type and a text-answerable flag.
5. Version groups for superseded decks and near-duplicates.
6. A redaction log covering pixels, text layer and file metadata.
7. A held-out split agreed before anyone tunes on the corpus.

For chunking and structure choices on the text side, see delivery formats that survive chunking; for embedded tables, see table and spreadsheet retrieval data.

Privacy and rights issues specific to page images

Page images leak information that text redaction misses, so de-identification has to cover pixels, the OCR or native text layer, and file metadata such as PPTX author fields and PDF XMP. A name blurred in the image but left in the text layer, or a logo in a slide master, will surface in retrieval. The pixel-level workflow is covered in redacting PII in scanned documents.

Check that redaction does not destroy the evidence a query depends on; masking a chart legend can make a chart-value query unanswerable. Rights matter as well: slides often embed third-party stock images, analyst charts and client logos that the supplying company may not own. Ask who owns embedded content and whether your license covers building and keeping multi-vector page indexes, which can be large [7]; see embedding and vector index rights.

How SourceX approaches visual document requests

SourceX sources operational datasets, including documents, from US companies on request and manages the commercial process, including the license. Nothing is held in stock, a request does not guarantee a match, and nothing is contracted until a supplier agrees. You describe the pages and queries you need, not the businesses; every release is approved by the supplying company.

Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Personal details such as names, emails, phone and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Delivery runs through private, access-controlled workflows after an executed agreement. If you are scoping a corpus now, you can send a visual retrieval data request to SourceX.

Source page-image retrieval data for your RAG system

Describe the page mix, query types and delivery format your visual retriever needs. SourceX looks for US businesses that hold matching slides, scans and PDFs, assesses data and licensing permissions, and agrees pricing and allowed uses in a license before anything is delivered. Start a buyer request.