All insights
XINDAR INSIGHT

Verify the PDF Page Behind an AI Citation

Resolve PDF page indices, physical page numbers and document labels with a public test file, a reproducible page-map script and a buyer acceptance procedure.

Sources and local examples checked November 17, 2026.

An AI citation to a PDF becomes reproducible when a reviewer can identify the file edition and the exact page under a stated numbering system. One sheet can have a zero-based parser index, a one-based position in the file and a PDF page label such as iv or A-8. The identifier printed on the sheet supplies another value to check. Before accepting “page 8” as evidence, establish what the number means and open the destination in the viewer your buyer uses. Then inspect the cited passage with its product, units and conditions.

Give each page identifier a defined meaning

Consider an engineering catalogue with introductory pages followed by a numbered technical section. The eighth sheet in the file could carry the printed number 4. A parser can address that sheet at index 7. A viewer with page-label support can present it as page 4.

Three fields make those references explicit.

FieldMeaningExample for the same sheet
Parser page indexThe position used by a specified zero-based programming interface7
Physical page numberThe sheet's one-based position in the file8
PDF page labelThe document's navigation label, where supplied4

The page's visible heading, footer or drawing-sheet identifier deserves its own recorded value. A document can display Sheet 4 of 12, use a navigation label A-4, and occupy physical page 8 after front matter. Each value helps a reviewer locate the intended evidence.

PDF supports page-label sequences with numbering styles and prefixes. The PDF Association describes the distinction between page index, page number and page label, including documents whose front matter and body use different sequences. Its examples also show why a document-wide offset can become inadequate when numbering changes again. PDF Association page-label guide

Read the fragment on the actual exported URL

RFC 8118 defines the PDF fragment parameter page=<pageNum> in terms of physical pages, with the first page numbered 1. Under that syntax, #page=8 identifies the eighth physical page. The numbering on the sheet can be different. RFC 8118, section 3

Suppose a fictional supplier manual begins with four Roman-numbered sheets. Its body then starts at printed page 1. The sheet showing printed page 4 occupies physical page 8. A reference that combines the manual's revision, #page=8 and the visible identifier “4” gives the reviewer a consistent starting point.

Record the link exactly as the monitoring system exports it. A URL ending in the PDF filename offers a different navigation instruction from one containing #page=8. An application might instead link to its own document viewer with a page query parameter. That application's parameter needs documentation or a direct test to establish its numbering convention.

Open the export in the actual receiving workflow. A browser may display the PDF, download it, or hand it to another application. Retain the observed destination and viewer details with the result. A procurement reviewer using a downloaded document needs a usable locator even when the original interface supplied an interactive source preview.

A public test file makes the mapping inspectable

The PDF Association supplies a sixteen-page page-label test document. In a local read with PyMuPDF 1.28.2, its first eight page labels were i, ii, iii, iv, 1, 2, 3 and 4. Selected rendered pages visibly state their page, index and label mapping.

Physical page in the test fileZero-based indexLabel returned for that page
10i
541
874

These observations establish an identifier mapping in a public compatibility fixture. A supplier's catalogue needs its own file and page check. Account acceptance can use the same workflow with a retained monitoring answer and its exported link. Public page-label test PDF

Export a page map before resolving the citation

A small local inventory can make the page positions and parser-returned labels available for review. The following Python program reads a PDF, records its SHA-256 fingerprint and emits one row per physical page. It was run on the public fixture using PyMuPDF 1.28.2.

Save the program as pdf_page_map.py in an environment with PyMuPDF available.

from pathlib import Path
import hashlib
import json
import sys

import pymupdf

pdf_path = Path(sys.argv[1])
file_hash = hashlib.sha256(pdf_path.read_bytes()).hexdigest()
page_map = []

with pymupdf.open(pdf_path) as document:
    for page in document:
        page_map.append({
            "page_index_0based": page.number,
            "physical_page_number": page.number + 1,
            "parser_page_label": page.get_label(),
        })

result = {
    "filename": pdf_path.name,
    "sha256": file_hash,
    "pymupdf_version": pymupdf.VersionBind,
    "pages": page_map,
}
print(json.dumps(result, ensure_ascii=False, indent=2))

To reproduce the public-fixture run, download the linked test file as PageLabelsTest.pdf and pass its path.

python pdf_page_map.py PageLabelsTest.pdf

PyMuPDF documents Page.get_label() as returning a label string, with an empty string when a label is undefined. The program retains the returned value. A reviewer can add the visible printed identifier after checking the page image. PyMuPDF page-label API

Use the fingerprint to identify the bytes inspected in this review. Record the document's title, publisher and revision alongside it. A matching fingerprint binds two copies of the file to the same bytes. Assessing the document's authority and the claim's support remains part of the evidence review.

Resolve missing and repeated labels explicitly

A label can occur on several pages. Compiled product manuals may restart numbering for each module, and prefixes or numbering styles may change between sections. Resolve the citation against its chapter, drawing number, table title or nearby passage when a label alone returns several candidates.

PyMuPDF's Document.get_page_numbers() searches for pages with a specified label and can return multiple matches. Its default search checks the document for all matching pages. That gives a reviewer a way to inspect an ambiguous label before choosing a destination. PyMuPDF label lookup

An empty label field also needs an interpretation recorded in the review. Keep the physical page number available, inspect the printed identifier, and document the evidence used to connect the two. An OCR reading of a footer can be retained with its extraction method and a comparison to the page image.

Preserve the original label when constructing a lookup key. Roman numerals, prefixes and other characters can carry the document's section structure. A normalized search value should travel with its transformation rule so another reviewer can reproduce the lookup.

Keep the file edition with the locator

Page positions can move when a supplier inserts a cover, adds a declaration or replaces an appendix. A citation recorded against one edition therefore needs enough information to recover that edition.

A useful evidence record contains the document title and revision, retrieved URL, retrieval time, file fingerprint, physical page number, parser-returned label, visible sheet identifier and cited passage. Add a table or figure identifier when it helps distinguish the relevant content on the sheet. These fields support a reviewer answering a concrete question, such as which pump curve governs the quoted operating condition.

Named destinations provide another navigation mechanism. Adobe's pdfmark documentation describes names that identify locations within a PDF and can be used by links or bookmarks. A publisher maintaining those destination names through revisions can give referring documents a more stable navigation target. Include the destination name in the evidence record and test the revised file's destination after publication. Adobe destination documentation

For a manufacturer publishing an English capability page, put the document revision and relevant section near the PDF link. A brief HTML explanation of the supported capability gives buyers context before they open the complete technical record. The explanation should retain the operating conditions and approval status that govern the claim.

A five-step acceptance check

Use one consequential citation from a retained AI answer and follow it through the receiving workflow.

  1. Identify the source file. Save the exported URL, document title, revision, retrieval time and fingerprint. Confirm that the inspection uses the intended edition.
  2. Name the locator convention. Establish whether the provided number represents a zero-based index, a one-based physical position, a page label or a printed sheet identifier. Record the source of that interpretation.
  3. Open the destination. Use the buyer's intended browser or PDF application. Record the visible page and any change of destination through the export or download workflow.
  4. Inspect the passage. Compare the claim with the page text, table headings, units, product variant and qualifying notes. Save the section or figure identifier that helps another reader repeat the check.
  5. Replay the record. Have a reviewer use the saved file identity and locator to reach the same evidence. Record ambiguous labels, missing files and changed editions with their actual outcomes.

The acceptance record can use a compact decision table.

Observed conditionReview actionEvidence to retain
Physical page and label map to the intended sheetInspect the cited passageFile fingerprint and page mapping
Several pages share the labelUse section and passage details to resolve candidatesCandidate positions and the selection reason
The link opens the correct file at its beginningTest a usable page or section locatorActual receiving-viewer result
The current file's fingerprint differsLocate the cited edition or review the changed fileBoth version identities and their locator results
The located passage has different product or test conditionsReassess the claim against those conditionsPage excerpt, relevant notes and review outcome

For a B2B team evaluating Semvi by Xindar, include one PDF citation in the full-answer and source-URL export demonstration. Ask to follow a retained answer through the exported link, file identity and reviewed passage. Verify the fields and receiving-viewer behaviour in the account you plan to use. Manufacturers can use the same record when preparing buyer-facing technical pages. Semvi product scope · Manufacturing content framework

Questions buyers ask

Which number belongs in a #page= PDF fragment?

RFC 8118 specifies the one-based physical page number. Resolve the printed identifier and PDF label against the actual file, then test the exported link in the receiving viewer.

Does a PDF label establish what is printed on the sheet?

Inspect both representations. W3C's PDF17 technique describes checking that navigation numbering matches the numbering visible on the document's pages. Its procedure combines visual checks with inspection of the page-label data. W3C PDF17 technique

What should a report do when the label is missing?

Retain the physical position and the parser's returned label value, then record the visible sheet identifier from a page check. Include a section, figure or passage locator that makes the evidence recoverable.

Can a file fingerprint establish that a statement is supported?

A fingerprint identifies the inspected file bytes. Support for the statement comes from reviewing the relevant passage, its conditions and the document's authority.

What should trigger another locator check?

Review the locator after a source-file revision, a changed exported URL or a change to the receiving viewer. Keep the previous record available so the reviewer can explain the new result against the earlier file and navigation conditions.

Source and method note

This article combines primary PDF documentation with a local read of the PDF Association's public test file. The page-map program was executed with PyMuPDF 1.28.2. Selected physical pages 1, 5, 8, 9 and 13 were rendered and inspected. The numerical supplier-manual example is fictional. A commercial account export and a supplier's own file each require their actual evidence and acceptance record.

Back to insightsMarkdown version