GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-08
An old door schedule can be readable on paper and still become unreliable when converted into searchable data. A number may survive extraction while losing the column heading or footnote that explains it. Docling is useful to examine at this boundary: the moment an architectural document becomes structured information that another person or system will trust.
The Docling technical report, first submitted on 19 August 2024, describes an open-source document-conversion package using specialised AI for layout and table structure. The official repository documents a released software toolkit with a unified document representation and export options. This review, dated 8 October 2026, proposes an architectural evaluation of those capabilities. It does not report a conversion test or announce the toolkit's original launch date.
A suitable first task is to recover an existing building's door schedule for a refurbishment archive. The output would help a surveyor locate references and prepare questions for a site visit. Its usefulness depends on retaining relationships: which width belongs to which door, whether a dimension describes an opening or leaf, and whether a note applies to one row or an entire group. These are editorial validation priorities rather than automatic guarantees supplied by the software.
Start with a controlled set of documents containing a native PDF, a scanned copy and a page with a more difficult table layout. Keep each original intact and record drawing number, revision, page and the provenance of the file. Choose the parser version and OCR settings explicitly. A successful run on a clean digital schedule should not be used to imply that a faded scan or a rotated drawing has the same reliability.
For a concrete pilot, manually transcribe twenty selected schedule rows before viewing the automated output. Include repeated door marks on different floors, an empty cell, a merged header and a note that qualifies a dimension. Treat each record as a linked set of values and their headings. Checking only whether the extracted words appear somewhere in the output would miss a width moved into the height column.
Compare the conversion cell by cell against that reference. Count omissions, unwanted additions and misplaced values separately. Review units, decimal points and symbols in the original page image. Where the source is ambiguous, preserve that ambiguity with a review flag instead of filling the gap from a nearby row. The test should allow a record to remain unresolved without making the entire archive appear complete.
Then ask a second reviewer to retrieve three specific records using only the converted archive and its source links. Can they return to the exact page and establish which revision supplied the value? A useful acceptance criterion is that every field intended for downstream use has a traceable source and an agreed review state. The criteria should be fixed before the trial; twenty rows alone do not establish an error rate for the whole archive.
Finally, keep extraction separate from the verified survey. A correctly recovered historic dimension may no longer describe the building. An updated PDF may also supersede an earlier schedule without replacing every drawing in the set. The architectural benefit comes from a better index into project evidence, with uncertain cells, revision conflicts and field checks still visible. Only reviewed records should move into a working schedule, and that transfer should preserve the document references that made the information checkable.