A file name is an address with no destination
“Source: report.pdf” identifies a container. It does not identify the passage that supports the claim.
The problem is easy to see in a long PDF. A report may contain an executive summary, a table, an appendix, and several qualifications. A claim can be somewhere inside all of that. The file name tells you none of it.
A page number is better. It narrows the search. It still may not be enough. A page can contain several columns, footnotes, charts, and captions. The reader still has to decide which words matter.
A citation should point to the words and their position on the page. That means a page number plus a bounding box. It is a small change in format. It creates a different object to work with.
What the ingest step keeps
When a source enters Doculate, the uploaded file is preserved unchanged. The system does not rewrite the PDF to make it easier to search. It does not replace the original with extracted text.
Parsing records located passages alongside that original. Each one can carry:
- the text that was found
- the page number
- a bounding box on that page
- character offsets in the parsed text
- a stable key for the span
The text makes the passage searchable. The page and box make it locatable. The offsets connect the passage to the parsed representation without losing its relationship to the source.
These are separate jobs. Search needs text. A reader checking a citation needs location. A review interface needs both.
| Citation | Settles | The reader still has to |
|---|---|---|
| File name | Which document | Read the document and find the passage |
| File and page | Which page | Scan the layout and pick the passage |
| File, page and region | Which words, with the label and qualification around them | Judge whether the passage supports the claim |
The same distinction applies when several small fragments form one useful passage. A single parsed line is often too narrow to carry a statutory subsection or a full paragraph. Neighbouring fragments can be grouped under their heading while each one keeps its own location. The group gives the passage enough context to read. The parts underneath keep the exact pages and regions.
Why the rectangle matters
Imagine a claim about an amount in a financial report. The number may appear in a table. The row label is on the left. The unit is in a heading above. A note below limits what the number means.
The supporting material is not always one clean sentence. It is a region of a page with structure around it. A rectangle gives the reviewer a concrete place to inspect. It can include the label, value, and nearby qualification.
The rectangle is also useful when text extraction is imperfect. A parser can return the right characters in an unexpected order. The page image remains the visual reference. The citation can lead back to the source area rather than asking the reviewer to trust a reconstructed document.
This does not mean every box proves every interpretation. Location is not reasoning. It is an honest boundary around the material that needs checking. The claim still has to fit what the passage says.
The original is part of the evidence chain
Preserving the upload unchanged is more than a storage preference. It separates the source from the representations derived from it.
The parsed text may be indexed. Fragments may be grouped into passages. Claims may cite those passages. None of those operations should alter the thing being cited.
That gives the evidence chain a fixed starting point:
- the original file is retained
- parsing records text and location metadata
- retrieval groups related spans without discarding their references
- a citation resolves through those references to a page region
Each layer has a purpose. Each layer can be inspected. If a parsed passage looks wrong, the original remains available for comparison.
The same separation prevents a familiar failure mode. Someone can clean up extracted text, remove a footnote, or normalise a number for display. If that edited representation becomes the only source, the citation points to something that no longer matches the uploaded material. An unchanged original gives the system a durable reference.
Citation is a resolution problem
A document-level citation asks the reader to perform the resolution. They must open the file, locate the page, scan the layout, and decide which passage supports the claim.
A rectangle-level citation records the resolution. It says which asset, which page, and which region the claim is tied to. It leaves interpretation open to review, but it does not hide the starting point.
That is the useful standard for evidence-backed documents. A citation should not merely name where a source lives. It should show where the support lives inside the source.