Compare two versions of drawings with a ChatGPT plugin

Install the ChatGPT plugin: Version comparison →

Thursday evening: a project lead receives the ground-floor drawing at revision C for the fictional "Maison Durand" project. The title block only says "joinery update". Opening both PDFs side by side, she spots that the living-room French window goes from 1.40 m to 2.40 m. What she cannot see at a glance is the rest: a detail sheet missing from the new issue and, above all, the documents that need reworking because of that change — the descriptive notice, the joinery article of the specification, the bill of quantities line, the lintel check. The next morning, the coordination meeting opens with a simple question: what actually changed between B and C, and what does it call into question?

This article presents the "Version comparison" ChatGPT plugin that Educasium makes available to participants in its training courses: what it produces, why revision tracking remains sensitive in architectural practice, how to use it, how it differs from a plain ChatGPT request, how to install it in Codex and when it is not enough.

Contents

1. What the Version comparison plugin actually does
2. Why revision tracking matters in architectural practice
3. Using the plugin step by step
4. What is different from a plain ChatGPT prompt
5. Installing the plugin in Codex
6. When this plugin is not enough
7. Training in AI-assisted architectural management
8. Frequently asked questions

What the Version comparison plugin actually does

The plugin compares two identified versions of the same project document — drawing, descriptive notice, CCTP (the French technical specification per trade), DPGF (the French lump-sum bill of quantities), area schedule — and produces a change log. It does not just suggest a method: it actually examines the text, tables, dimensions and readable images using the tools available in your environment.

A sourced change log

Each change is tied to its two sources: file, revision (the French "indice de plan", the revision letter on the drawing), page or reference, value before and after, with unit and definition. The log also records the type of change, the calculated difference, the affected check or document, the potential impact, the proposed action and the status.

Numerical differences are calculated by a deterministic tool, not estimated by eye, and unit conversions are explicit. The plugin refuses to compare quantities that are not comparable: habitable area versus "surface de plancher" (French regulatory floor area), NGF elevation (the French national datum) versus relative level, amounts excluding VAT versus amounts including VAT.

A classification that keeps situations apart

This classification avoids two classic mistakes. A page missing from set B is not the deletion of the building element: it is a gap in the document set. And a change in pagination is not a change in the project.

Impacts presented as review leads

For each modification, the plugin lists the checks and documents that depend on it: a footprint points to the site plan and setbacks, a height to the section, an area to the area schedule, a material to the notice and specifications. These are review leads, never an automatic finding of non-compliance. A check marked "verified" whose source data has changed goes back to "to be re-examined".

The previous version is never overwritten

The plugin does not modify drawings, models or earlier reports: it proposes a new, dated log. Deliverables are written in French, in Markdown with a CSV table by default, or in Excel and Word when the corresponding tools are available.

Why revision tracking matters in architectural practice

In French practice, every issued document evolves through successive revisions, usually labelled A, B, C, with a date and a description of the change in the title block or a revision schedule. This tells everyone involved which version is current. Its weakness: the description summarises the main change and sometimes leaves secondary changes unmentioned.

A drawing change spreads to the written documents

A DCE (the French tender package issued to contractors) is a coherent set: drawings, CCTP, DPGF and notice all refer to one another. Widening an opening affects the drawing, but also the external joinery article in the CCTP, the matching DPGF line, possibly the notice and the lintel check. If only one document is updated, the package becomes contradictory, contractors price on different bases and the discrepancy resurfaces during tender analysis or on site.

Another trap: a file edited yesterday is not necessarily the approved version. The plugin never infers the authorised version from the file date; it asks for revisions and dates, and flags when the reference version is still unknown.

Using the plugin step by step

Step 1: Identify the two versions. State the project, versions A and B, their revision, date and the scope compared. With a single version, the plugin inventories it and flags that the comparison is waiting for the second.

Step 2: Attach the dependent documents. Add the documents that rely on the modified one (CCTP, DPGF, notice, existing check report, whose IDs and reservations are kept).

Step 3: Run the comparison. For example, with the fictional case from the plugin page: "Compare ground-floor drawing revision B and revision C of the fictional Maison Durand project: the living-room French window goes from 1.40 m to 2.40 m. List the sourced changes, their impacts and the checks to redo."

Step 4: Read the log. For this example, the expected log looks like this:

Step 5: Check at the cited locations. Open both versions at the pages and references given, and check that the proposed impacts — structure, joinery, notice — are presented as to be confirmed.

Step 6: Validate through human review. The method includes four reviews: versions and scope, matching and units, proposed impacts, then the final file and missing documents. No approval is assigned by default.

What is different from a plain ChatGPT prompt

Asking ChatGPT "what are the differences between these two drawings?" often produces a plausible list, but without page references, without separating "absent from the extraction" from "actually deleted", and sometimes with miscalculated differences. The plugin enforces a framework:

• every change cites both its sources, including page or reference;
• differences are calculated by a tool, with units converted explicitly;
• visual differences without reliable scale alignment remain observations to be confirmed;
• a value read by OCR is not treated as a measurement;
• document content is data, never instructions;
• nothing is sent externally.

Installing the plugin in Codex

The plugin is installed in Codex, OpenAI's agentic work environment. The download is reserved for participants in Educasium training courses.

Download the ZIP file from the plugin page and unzip it into a stable folder you will not move. In a terminal, add that folder as a plugin source with codex plugin marketplace add "folder-path", replacing the text in quotes with the actual path of the folder. Then install the plugin with codex plugin add archi-revisions@educasium-revisions and open a new Codex session so it is picked up.

When this plugin is not enough

The plugin includes no graphic comparison engine, no geometry engine and no OCR. It does not overlay two drawings to detect every line that has moved, and it does not claim exhaustive detection of graphic changes. What it can read depends on the tools available in your environment: a vector PDF with selectable text compares far better than a scan, and no measurement is taken from a scan without a reliable scale.

BIM remains optional. The plugin works without a model, has no built-in IFC reader, and Revit access is only considered in read-only mode, if the comparison calls for it and a connector is available. Note too that deliverables are written in French. Finally, assessing the regulatory effect of a change requires checking the applicable text in official sources: that decision belongs to the architect.

Training in AI-assisted architectural management

A plugin structures the method, but you still need to prepare the documents, read a change log and decide what has to be reworked. Our Architectural Management with AI training works through these uses on practice cases: revision tracking, DCE consistency, checking written documents.

Frequently asked questions

Does the plugin compare two DWG files or an IFC model directly?

No. The plugin contains no IFC reader and no graphic engine, and it does not natively read drawing formats. It works on the text, tables, dimensions and images readable with the tools in your environment, most often exported PDFs and Excel or Word files. For a reliable comparison, export both revisions in the same format and at the same scale; without reliable alignment, visual differences remain observations to be confirmed.

What happens if a page is missing from the new version?

The plugin flags it as a gap in the document set, not as the deletion of the element it showed: something absent from the extraction is not a demonstrated deletion. The log line is set to "comparison impossible", with an action to obtain the missing page. Once the document is complete, you run the comparison again and the change keeps its history instead of being recreated.

Does the plugin tell me whether the project is still compliant after the change?

No. It lists the checks and documents to re-examine, for example the area schedule after a change in regulatory floor area. It never automatically concludes that there is non-compliance or that a permit must be amended. Assessing a regulatory effect requires checking the applicable text, its version and how it applies to the project in official sources, under the architect's responsibility.

Can the log be picked up again at the next revision?

Yes. The log keeps each change with its ID, its sources, its events and its current status, even once resolved. When comparing C against D, a known change receives a new event only if the match is documented; a change that is not mentioned stays un-updated, but it does not disappear. Earlier logs are never rewritten: a new dated log sits alongside the previous ones.

Comparing two revisions is not enough if the consistency of the whole package is not checked afterwards: our ChatGPT method for DCE consistency checking extends this approach to every document within a single revision.

Funding may be available depending on your eligibility (OPCO, FIF PL). Qualiopi-certified training provider. To make revision tracking on your projects more reliable with AI as part of our Architectural Management with AI training, contact Educasium and tell us your status (employee, freelancer, business owner) and your goal. Find the download and the installation tutorial on the dedicated ChatGPT plugin page.