How to turn a practice's archive of past projects into an AI knowledge base, without mixing up projects

An architecture practice or engineering firm with years of work behind it has already solved much of what each new commission asks for: the construction description of a similar building, the building-code compliance justification for a use it has designed before, a façade detail or the specifications for a type of work. The problem is finding it: it is spread across folders by year, in PDFs and in versions nobody remembers which was the final one.

With AI, that archive becomes a knowledge base you can query in plain language, with answers that cite the project, the document and the page. For it to be useful, three things need doing first: separate each project file, distinguish approved documents from drafts, and record which regulations each project was designed under.

In Spain, those approved documents are the "proyectos visados", stamped by the professional association. The method is the same in any country.

Which documents are useful

• Project descriptions: general, construction and calculation reports. These are what gets reused most.
• Building-code compliance justifications, document by document. In Spain, the Código Técnico de la Edificación (CTE) covers structural safety, fire safety, safety in use and accessibility, health, noise protection and energy saving.
• Specifications, bills of quantities and budgets, including BC3 files (the Spanish exchange format), which are read item by item.
• Site documentation: site visit reports, construction management reports, valuations and completion certificates.
• Drawings in PDF. Their text is used: title blocks, legends, written dimensions and notes. The geometry of the drawing is not interpreted as in a CAD program.
• Older scanned projects. They are read with text recognition, with more errors the worse the scan.

Step by step

1. Inventory. Which projects exist, where each one is and in what format. You do not need to digitise the whole archive on day one: start with the projects most similar to today's commissions.
2. One record per project. Each project goes in with its data: use, floor area, municipality, year, client and status. Searches can then filter ("multi-family housing in Madrid since 2020") and the system knows which project each page belongs to.
3. Approved documents separated from drafts. Only approved or signed documents go in as a reference. Drafts can be added, but flagged, so they are never mistaken for the final version.
4. The regulations for each project. A 2012 project was designed under different energy-saving requirements from today's: the Spanish CTE was approved in 2006 and its energy-saving document was updated in 2013 and 2019. The system records which version each project used and warns when an old solution no longer meets current regulations.
5. Answers with their source. Each answer cites the project, the document and the page, so it can be checked in one click. If something is not in the archive, it says so rather than inventing it.
6. Permissions. Each person sees the projects assigned to them, and projects with confidentiality clauses stay out of general search.

Without mixing up projects

It is the most serious mistake and the easiest to make: the AI taking the floor area of one project and putting it into another project's description. It is avoided by separating three things:

• Criteria and reusable solutions: how your practice handles a justification, a construction detail or the wording of a section. These are reused.
• Data exclusive to each project: owner, plot, areas and amounts. They never pass into another document.
• The regulations each project was designed under, to compare them with those in force.

That is how the AI document assistant for architecture and engineering (in Spanish) works: it reuses criteria and wording, never a client's data.

What you can ask it

• "How did we justify evacuation in the last car park we designed?"
• "Find a detail of a ventilated façade meeting a floor slab and tell me which projects we used it in."
• "Prepare the construction description for this house from the two most similar ones we have had approved."
• "Which waterproofing items did we use on flat roofs in 2023?"

The answers are drafts with their source. The technician decides what is used and signs the document.

What to avoid

• Uploading the archive to a general-purpose AI. Projects contain owners' and developers' data, and their design belongs to the practice. The archive should be in a private environment, with the data in the European Union and not used to train models (where the data is processed, in Spanish).
• Mixing drafts and approved versions without flagging them.
• Accepting an old solution without checking which regulations it was designed under.
• Expecting it to calculate or draw. It writes from the results of the practice's software; calculation and drawing stay there.

How to start and what it costs

Start with a sample: the projects most similar to current commissions and one family of documents, for example construction descriptions or code compliance justifications. With it you measure how much time each document saves, including review time, before loading the rest of the archive.

At Tecnea, the proof of concept starts at €850, full development ranges from €5,000 to €20,000 and the monthly fee starts at €250 (what each amount includes, in Spanish). More use cases in AI for architecture practices and for engineering firms (in Spanish).

Frequently asked questions

Do we need to digitise the whole archive?

No. Start with the most useful projects and add the rest gradually. Anything already in text-based PDF is used directly, and scanned documents are read with text recognition.

Does it understand drawings?

It reads their text (title blocks, legends, written dimensions and notes) and references them, but does not interpret geometry as a CAD program would.

Does it work for an engineering firm?

Yes. The method is the same with technical reports, calculation annexes, construction management reports and valuations.

What about projects designed under repealed regulations?

They are still useful as a reference for criteria. The system shows which version of the regulations they were designed under, so the technician can check what has changed.

If you would like to see how it would work with your archive, tell us how many projects your practice has and how they are stored.

This article is informational. Tecnea develops custom document assistants for architecture practices and engineering firms, so we have a commercial interest in the topic. The references to the Spanish CTE summarise its versions and do not replace checking the text in force.

Sources

1. Real Decreto 314/2006, de 17 de marzo, approving the Código Técnico de la Edificación: boe.es
2. Orden FOM/1635/2013, de 10 de septiembre, updating Basic Document DB-HE "Energy saving": boe.es
3. Código Técnico de la Edificación, "El CTE en el BOE" (including Real Decreto 732/2019): codigotecnico.org