Construction project data can support AI agents and models when it contains enough of a project's history to define a task and check the answer. Drawings alone can help an agent read a design. To ask what changed during delivery, the agent also needs records from other stages, with reliable dates and project links.
A project archive is a starting point, not a ready training set. Before preparing thousands of files, choose one task and inspect a few completed projects.
Start with a question an agent could answer
Consider a request such as: “Which design changes affected the equipment ordered for this project?” A person would look at the original drawing or specification, a later revision, the quote or purchase record, and perhaps an issue record explaining the change. The answer needs a cited trail through those files.
For that task, a useful record might contain:
| Record | What it could contribute | What needs checking |
|---|---|---|
| Drawing or design document | The intended scope and its revisions | Is the revision date and project ID reliable? |
| Quote or purchasing record | What was priced or ordered | Can line items be tied to the same equipment or project? |
| Work order or issue record | What happened during delivery | Does it describe the change, or only mention a problem? |
| Commissioning or handover record | The final state | Is it complete enough to check the proposed answer? |
This is an illustrative task design, not a claim that any archive already has these joins or labels. If a link has to be guessed from a filename, record that uncertainty. If the final state is missing, the task may work as document retrieval but not as an evaluation with a reliable answer.
Test one project before building a dataset
Select a small set of completed projects with different sizes and dates. For each one, make a simple timeline. List the relevant files, their source paths, project IDs, revisions, and the decision or outcome they appear to document. Then ask a domain expert to answer the chosen question from the records alone.
The pilot should tell you three things: whether the input records exist, whether their links survive review, and whether the outcome can be checked without relying on private knowledge from the project team. Keep a separate status for missing or disputed links. Do not turn those gaps into confident labels.
Research benchmarks show why the check matters. WorkArena defines work tasks with a way to validate an agent's result. A historical project archive needs comparable task and checking rules to support an agent evaluation. Its raw files do not supply those rules automatically.
An engineering archive to inspect
Orla's SolarEPC-1TB is an available solar project archive covering work from early site studies and design through construction, commissioning, and handover. Its published card lists drawings, reports, quotes, purchasing, work orders, issue records, correspondence, and handover material. The roughly 1.2 TB, 240 projects, and file counts on the card are estimates. The archive is primarily in Portuguese.
The downloadable sample is a selection of technical project records, not a representative workflow sample or a prepared agent benchmark. A team interested in the example task should ask Orla for project coverage, linked record availability, and a scoped sample before assuming the necessary joins exist.
For the general record structure behind this exercise, read what makes an enterprise workflow dataset useful for AI.