Uploading a PDF does not establish that an AI system received every relevant page, table and note. Content can disappear during recognition, conversion to text or transfer to the next component. A convincing summary does not reveal what was left out.
A useful small-business control belongs before search and summarisation: compare the original with the representation the system will actually use, such as extracted text or page images. The workflow below is an editorial proposal, not a certified standard or a report of product testing.
Who needs this check
This guide is for operations managers, purchasing staff and support teams working with catalogues, instructions and supplier document packs. Its purpose is to catch missing material before that omission becomes an incorrect customer answer.
Separate receiving a file, extracting its contents and using the extracted material. Evidence of the first stage does not establish the others. Even a complete representation might later be excluded from a particular answer by retrieval or context selection. This guide addresses the earlier question: is the required content present in the representation available to the system?
1. Record the original and the required scope
Keep the source file unchanged. Record its name, identifier, version and page count in a PDF viewer. If your system supports a checksum, include it to distinguish identically named copies. Keep any working version processed with optical character recognition (OCR) separately.
Define what the task needs: body text, tables, footnotes, diagrams, handwritten annotations and appendices. A reference to another file does not mean that file has been received. List any mentioned but missing attachment as a gap.
Count physical pages from the beginning of the file and record printed page labels separately. Covers, contents pages and Roman numerals can shift the relationship. An explicitly agreed exclusion of a promotional cover is reasonable; silently excluding delivery conditions introduces a different risk.
2. Create a page coverage register
Open the original beside the extraction output. Make a short record for each page:
- Physical page number and printed label.
- Expected content: text, table, image, note or a genuinely blank page.
- Processing method: text layer, OCR, image inspection or manual transcription.
- Reference to the corresponding passage or image in the working copy.
- Status: checked, review required, missing or exclusion agreed.
- Issue, responsible person and review date.
Build this register from the original, not just the extractor’s inventory. Otherwise, a missing page might also be absent from the list being checked. Do not assign “checked” automatically because some text exists: the only recognised content could be a page header.
On each page, compare the arrangement of major blocks and the beginning and end of the content. Inspect decision-critical regions separately. Investigate unexpectedly sparse text: the page might be blank, contain a diagram or be an unsuccessful scan. Character counts can locate suspicious areas but do not measure content completeness.
3. Distinguish text layers from scans
A single PDF can combine digital text and scanned pages. The pypdf documentation explains that it extracts text but does not recognise text inside images. A scan can also contain an existing OCR layer with errors.
Try selecting and copying a passage from a difficult page, then compare it with the visible text. This is a diagnostic clue, not proof that the whole document is correct. Use an approved OCR tool or another supported image-reading route for image-only pages. Check languages, orientation and the legibility of small markings.
One specific trap is documented by OCRmyPDF: its companion sidecar text file excludes pre-existing page text and pages skipped during OCR. These are omissions from the sidecar, not a statement that pages were deleted from the resulting PDF. If the workflow requires the full text, extract it from the resulting PDF with page references and check that output.
Do not convert every page to an image merely to obtain a uniform format. First identify which pages need additional processing, then inspect the consequences of conversion in a working copy.
4. Check relationships inside tables
A list of correctly recognised numbers does not establish which products, units or conditions they belong to. For each required table, compare column labels, row order, merged cells, units, continuation across pages and associated notes. Do not replace an empty cell with zero or guess a missing unit.
This concern has documented examples: Amazon Textract identifies possible inconsistent extraction with merged cells spanning multiple columns and irregular table structures. The review proposed here should identify discrepancies in your material, rather than assess the service as a whole.
Inspect the surrounding text too. The Azure Document Intelligence v4.0 documentation states that table and figure bounding regions exclude associated captions and footnotes. A cropped table image alone may therefore fail to preserve its qualifications.
5. Check the material actually passed onward
A verified extraction file can differ from what the next component loads or indexes. Where available, inspect the actual export, page inventory and processing completion record. Preserve the mapping between the original, extraction output and version passed to the system.
For an important item, record its page, exact passage or region coordinates, reading method, working-copy reference and verification result. This distinguishes three states: present in the original, retained in the representation and passed onward. Do not replace missing evidence with the model’s statement that it read the whole file.
If the application hides intermediate output, record that limitation. Manually comparing required sections with an answer can help complete a narrow task, but cannot establish the completeness of a hidden internal representation.
A fictional supplier-pack example
An imaginary shop receives a 12-page PDF: a cover, specifications on pages 2–7, a table on 8–9, conditions on 10 and scanned appendices on 11–12. In this teaching scenario, extraction creates records for every page, but the appendices lack readable text. A note beneath the table is also absent from its working copy. This is a fictional situation, not an observed test result.
The main text states a standard lead time of five working days. Scanned page 11 specifies 15 working days for a non-standard colour. The note on page 9 says orders must be placed in whole boxes, while page 12 explains that each box contains 12 units.
A summary of only the digital text is insufficient for answering the non-standard-colour question. The responsible person should hold that conclusion for review, run OCR on the appendices, reconnect the note to its table and compare the lead time and ordering unit with the original. Only then should the corrected representation be passed onward for drafting an answer. If required content is absent from the material passed onward, changing the model’s prompt will not restore it.
When to stop processing
Stop using the document for the affected decision if a required page is absent, important material is unreadable, a table has lost its note, or transfer of the checked copy cannot be confirmed. Request a better file, use an approved alternative extraction method or have a person transcribe the required content with source references.
Allow work to continue when required pages are accounted for, critical elements are checked, exclusions are explicit and unresolved questions remain visible. Sampling does not prove every character correct; representation completeness establishes neither the truth of the original nor a correct AI answer.
Availability and limits
OCR methods, image support, exports and page restrictions depend on the tool, version and plan. Check the actual processing route; there is no universal “PDF for AI” limit. This workflow can use an ordinary register, but access to intermediate results determines what can be verified. Process confidential files only in locations approved by your organisation.
Sources checked on 24 October 2026. Prepared with AI assistance (OpenAI); the workflow and example are editorial proposals, not a tested implementation.
