Construction project records for AI training

A construction project dataset is the record a building project leaves behind: requests for information and the design team's answers, submittals and review actions, change orders with cost and schedule impact, daily logs, inspections, punch lists, site photos and schedule updates. SourceX sources these records from established contractors and construction managers running Procore, Autodesk Construction Cloud or Oracle Aconex, removes personal data and location details, and reviews design-team and owner rights before licensing.

Request this dataset Email the data team

Dataset manifest

Sourced to your spec

What it is: RFIs, submittals, change orders, daily logs, inspections, photos and schedules from building projects
Typical systems: Procore, Autodesk Construction Cloud, Oracle Aconex, Primavera P6, Bluebeam
Typical history: Project records spanning years of operations; varies by partner
Modality: PDF documents, site photos, structured logs and native schedule files
Delivery formats: Agreed per order; JSONL records with original PDFs, photos and schedule files
Preparation: Names, addresses, faces, plates and photo GPS removed; payroll and injury records excluded
Licensing: Scope and permitted use agreed per order; drawings need design-team authorization
Availability: Sourced to your spec from partner contractors; not guaranteed

Request this dataset

What a delivery contains

Fields vary by source system and are fixed per order. A typical delivery includes:

Field	Type	What it holds
record_id	string	Pseudonymous ID, stable across the delivery so RFIs, change orders, photos and logs can be joined.
record_type	enum	RFI, submittal, change event, change order, daily log, inspection, punch item, meeting minutes or schedule update.
project	object	Pseudonymous project ID with building type, delivery method, contract type, region and the phase the records cover.
dates	object	Created, due, responded and closed timestamps, from which turnaround and days open can be computed.
ball_in_court	array	The roles that held the item in turn — subcontractor, general contractor, architect, engineer, owner — with hand-off times.
references	object	MasterFormat spec sections, drawing sheet numbers with revisions, and location on site by level, area or grid line.
text	string	The question, scope description, log narrative or inspection observation as the project team wrote it.
response	object	The design team's answer to an RFI, or a submittal review action such as approved as noted or revise and resubmit.
impact	object	Cost and schedule impact, reason code such as owner request, design error or unforeseen condition, and approval status.
daily_log	object	Weather, crew counts by trade, equipment, deliveries, delays and inspections for the day, without worker names.
photos	array	Site photo IDs with capture time, location on site and the record they document; faces, plates and GPS removed.
schedule	object	Activity IDs with baseline, forecast and actual dates from native schedule updates, linked to the records that moved them.
attachments	array	Submittal packages, sketches, inspection reports and meeting minutes as PDFs, where cleared for delivery.

Example record

{
  "record_id": "rfi_0217",
  "record_type": "rfi",
  "project": { "project_id": "prj_3b8e", "building_type": "outpatient_clinic",
               "delivery_method": "cm_at_risk", "region": "[REGION]", "phase": "construction" },
  "subject": "Supply duct conflicts with beam at grid C/4, level 2",
  "created_by_role": "mechanical_subcontractor",
  "dates": { "created": "2023-09-12T15:04:00Z", "due": "2023-09-19",
             "closed": "2023-09-20T13:12:00Z" },
  "ball_in_court": [
    { "role": "general_contractor", "from": "2023-09-12T15:04:00Z" },
    { "role": "architect", "from": "2023-09-13T10:20:00Z" },
    { "role": "mep_engineer", "from": "2023-09-14T08:02:00Z" },
    { "role": "general_contractor", "from": "2023-09-19T18:30:00Z" }
  ],
  "references": { "spec_section": "23 31 13", "sheets": ["M-201 rev 3", "S-202 rev 2"],
                  "location": "L2, grid C/4" },
  "text": "24x14 supply duct hits bottom flange of beam at C/4. Can we go to 30x10, or drop below?",
  "proposed_solution": "Resize to 30x10 to hold ceiling height.",
  "response": { "by_role": "mep_engineer", "attachments": ["sk_m_014.pdf"],
                "text": "Use 32x10. Keep 2 in. clear of sprinkler main. See SK-M-014." },
  "impact": { "cost": true, "schedule": false, "reason": "design_coordination" },
  "linked": {
    "change_event": "ce_041",
    "change_order": { "co_id": "co_012", "amount": 4820.00, "currency": "USD",
                      "days": 0, "status": "executed" },
    "photos": [ { "photo_id": "ph_8812", "taken_at": "2023-09-12T14:51:00Z",
                  "location": "L2, grid C/4", "faces_blurred": 1, "gps_removed": true } ],
    "daily_log": "dl_prj_3b8e_2023-09-12",
    "schedule_activity": "A2140"
  },
  "days_open": 8
}

Synthetic record for illustration. Field names, structure and format are agreed per order.

What AI teams use it for

Train RFI and submittal assistants

Real RFIs paired with the design team's answers, and submittals paired with review actions, show how questions get framed, routed and resolved against the drawings and specifications.

Forecast change-order and schedule risk

Records that link coordination problems to change orders and schedule activities show which issues end up costing money or time, and which close without either.

Ground retrieval over project documents

Specs, submittals, meeting minutes and RFI logs support questions such as which substitution was approved or why a detail changed, with recorded answers to test against.

Train vision models on located site photos

Photos tied to a date, a location on site and the record that prompted them give context for progress tracking, punch-item detection and safety models.

Evaluate construction agents on closed items

Closed RFIs and change orders with recorded outcomes become private test items: does the agent route, answer or price the issue the way the project team did?

Use-case guides: Document AI, enterprise search and RAG, Enterprise and computer-use agents, Robotics and embodied AI

What makes this data valuable

Linked chains

RFIs that point to the change events, change orders and schedule activities they caused show cause and consequence.

Ball-in-court history

Hand-off timestamps between subcontractor, contractor and design team show where items waited.

Spec and sheet references

MasterFormat sections and sheet numbers join each record to the part of the design it concerns.

Projects through closeout

Completed projects carry final change-order totals, punch completion and warranty items.

Schedule snapshots

Native schedule files saved at each update show how forecast dates moved, which a single PDF printout cannot.

Located photos

Photos tagged by level, area or grid line and linked to log entries carry far more context than unlabeled site folders.

The chain of records matters more than any single record

On a construction project, the links between records carry more knowledge than any record on its own. Take a duct that will not fit under a beam. The conflict surfaces in a clash report or a coordination meeting, becomes an RFI, is answered with a sketch, turns into a change event, a priced change order and a revised schedule activity, and finally a punch item if the fix was not finished cleanly. Each record alone is a fragment. The chain shows how a field problem was diagnosed, who decided, what it cost and how long it took.

Public sources rarely carry that chain. Building codes, specification templates and public bid documents are easy to find, and site-photo collections exist for narrow tasks such as hard-hat detection, but none connect a question raised on site to the decision that answered it and the money and days it consumed.

The chain is only as complete as the team's habits. Some projects run every RFI, submittal and change through the platform. Others use it for formal logs and settle the real discussion in email, calls and site meetings. The same software can hold a rich history on one project and a thin one on the next, so completeness is judged project by project, not by platform.

Whose archive it is shapes what it holds

The firm that keeps a project archive determines which records it holds and which it can license. A general contractor or construction manager holds the coordinating record: RFIs from every trade, submittal logs, owner change orders, the master schedule and daily logs. A specialty subcontractor holds a narrower but deeper record of its own scope — shop drawings, fabrication and installation detail, its change requests and its side of disputes — which suits trade-specific models. Owners and their representatives hold budgets, pay applications and a view across many projects, but less of the field detail.

For that reason a dataset is assembled project by project rather than by pulling RFIs at random, which would break the chains. Completed projects carry outcomes active ones cannot: final cost and schedule, punch completion, closeout and early warranty calls. Active projects add current practice and tools but leave those outcomes open, so a mix is common when recent work matters. The manifest for each candidate dataset lists the projects included, the phase each reached and what was removed before delivery.

What to check before licensing

Separate what the contractor authored, such as RFIs, logs and photos, from design documents. Drawings and specifications usually remain the design team's property, so they need its authorization or are excluded.
Ask how confidentiality clauses in owner contracts were reviewed, and whether projects with security requirements — government, defense, data centers, critical infrastructure — are excluded.
Check sample photos for faces, license plates, neighboring properties and, on residential work, the interiors of occupied homes. Confirm GPS and other EXIF metadata is stripped.
Confirm that prevailing-wage payroll reports, worker sign-in sheets and injury reports are excluded, or that anything kept from them is de-identified.
Measure linkage on the sample — the share of change orders that trace back to an RFI or change event, and how many schedule updates survive as separate files.
Check whether spec references follow the same MasterFormat edition across projects. Older projects may use the earlier 16-division numbering and need a mapping.
Compare building types, delivery methods, regions and project sizes against the distribution your model needs.

How licensing works through SourceX

1
Define
Send the domain, modality, volume, format, timeline and permitted use you need.
2
Source
SourceX identifies businesses that hold matching data and are open to licensing it.
3
Qualify
Fit, rights and quality are checked, and you review samples before committing.
4
License
Scope, permitted use, exclusivity, price and obligations are agreed in writing.
5
Deliver
Approved data is prepared, de-identified where required and transferred securely.

Questions buyers ask

Can I license real construction project records for AI training?

Yes, from contractors and construction managers that agree to license their project archives. SourceX matches your spec — building types, record types, regions and phases — against partner projects, then separates what each firm can license itself from design documents and owner-restricted material. Which firms hold matching projects and agree to take part determines supply, which is not guaranteed.

Who owns the drawings and specifications in a project archive?

Usually the architect and engineers. Under standard-form contracts common in the US, such as the AIA documents, the design team keeps copyright in its drawings and specifications, and the owner and contractor receive a limited license to use them for the project, unless the contract says otherwise. A contractor's archive therefore holds design documents it cannot license on its own, so they are excluded or cleared with the design team and owner.

How are workers and the public handled in site photos?

Faces, license plates and other identifying details are blurred, and GPS and other photo metadata are removed before delivery. Photos of occupied homes, neighboring properties or security features may be left out altogether. Daily logs usually report crews as headcounts by trade, but sign-in sheets and some notes name individuals, so those fields are removed or pseudonymized. The rules are agreed during scoping.

Are schedules delivered as native files?

Where the partner kept them, yes. Native Primavera P6 (XER) or Microsoft Project files preserve activities, logic links, baselines and progress, which PDF printouts lose. The most useful archives keep each periodic update, so you can see how forecast dates moved and match the movement to RFIs, change orders and delays noted in daily logs.

What happens to projects in a dispute or claim?

They are usually left out or delayed. Records from a project in litigation, arbitration or a pending claim may be under a legal hold or a settlement confidentiality term, and claim analyses prepared for lawyers can be privileged. These constraints are identified when the archive is scoped, and affected projects are excluded unless they can be cleared.

How are subcontractor bids and pricing handled?

As confidential commercial information. Subcontractor names are pseudonymized, and bid amounts, unit prices and markups can be removed, banded or kept as relative values, depending on what the partner's subcontracts allow and what your use needs. Change-order amounts are often the most useful cost signal, and whether they stay is agreed case by case.

Field service and maintenance work orders
Work orders tracing symptom, diagnosis, parts and fix, with asset histories and photos
Enterprise document archives
A company's working files with folders, versions and sharing metadata
Manufacturing quality and inspection records
Inspection results, nonconformances, dispositions and corrective actions from production plants
First-person video of skilled manual work
First-person video of skilled workers doing real tasks at partner businesses
Enterprise workflow and task execution histories
Linked task trajectories from request to outcome, across every tool the work touched

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Tell us what your models need

Send your spec — domain, volume, format, timeline and permitted use — and SourceX will match it against partner data and come back with what can be licensed.