Can your PDF extraction support the citation you want to make?
Finding the right words is only the first half of document retrieval. A useful citation also needs enough surviving structure to show where the words came from and what they belong to.
No PDF parser, OCR system, retrieval pipeline, model, or product was installed or run for this note. The failure modes are practical inspection cases, not observed results for OpenDataLoader PDF or any other tool.
Text presence is not citation support
A plain-text extraction can contain every visible word and still lose the relationships that make a statement checkable. Two columns may interleave. A table value may separate from its row and column labels. A footnote marker may survive while its footnote moves elsewhere. A caption may detach from its figure. A page number may be missing or may refer to the PDF file rather than the document's printed numbering.
Tagged PDF can carry intended reading order and semantic structures such as headings, lists, tables, and figures. Those structures help reuse and accessibility, but their presence in a file does not guarantee that every extractor preserves them or that the source tags are correct. An inspection therefore has to compare the extracted record with the document, not merely count characters.
Five failure modes worth checking
| Failure mode | What may look successful | What a reviewer should inspect |
|---|---|---|
| Reading order | All words are present. | Paragraphs follow the intended column, sidebar, and continuation order. |
| Table relationships | Cell values appear in the output. | Each value remains associated with the correct row header, column header, and table title. |
| Footnotes | Marker and note text both appear. | The marker links to the right note, and the note remains distinguishable from body text. |
| Figures and captions | Image and caption are separately extracted. | The caption stays attached to the intended figure, with page and figure identity preserved. |
| Citation anchors | A search returns a matching sentence. | The record retains a stable page or element anchor that another person can reopen and verify. |
Inspect from the claim backward
Start with the claim a reader might cite. Locate its exact supporting span in the PDF. Then ask whether the extracted record preserves the smallest context needed to interpret that span: heading, paragraph, table headers, caption, footnote, page, and element identity as applicable.
A citation-support record should distinguish at least three things:
- Source location: the PDF identity, page and visible region or semantic element.
- Extracted representation: the exact text and structure the pipeline retained.
- Review conclusion: whether that representation is sufficient for this particular claim, plus any ambiguity a person still needs to resolve.
This is claim-specific. An extraction may be adequate for citing a paragraph while being inadequate for citing a table on the same page.
Use the checklist before retrieval evaluation
The blank checklist prompts a reviewer to record source identity, the proposed claim, relevant structure, extraction evidence, and unresolved ambiguity. It contains no document content, parser output, product result, or approval.
Download the blank citation-support checklist →
What the primary sources establish
The PDF Association's Tagged PDF Q&A explains that tags can express intended reading order and semantic structures including headings, lists, tables, and figures. Its specific G4_03 multi-column technique demonstrates a two-column example whose logical tag order matches the author's intent.
The official OpenDataLoader PDF repository describes structured PDF extraction, and its JSON schema defines output fields for document metadata and nested extracted elements. Those sources establish a concrete structured-output surface worth inspecting. They do not establish how the product would perform on an unrun document or whether an extracted citation is correct.
Primary sources checked. The claim-backward inspection method, failure-mode table, and checklist are original synthesis.