For construction project data, retrieval-augmented generation (RAG) is usually the better fit than fine-tuning, because project records change every day, answers must point back to the exact document and revision, and access must follow project permissions. Fine-tuning has a role in shaping how a model writes or classifies content, but it is a poor way to teach a model the facts held in your common data environments (CDEs).
Two ways to make AI use your content
Fine-tuning adjusts a model's internal weights by training it on examples. After training, the model has absorbed patterns from that material and responds in a way that reflects them. RAG leaves the model unchanged. Instead, each time someone asks a question, the system searches your content, retrieves the most relevant passages and gives them to the model as context. The model writes an answer grounded in those passages and can cite them. Our article How Does RAG Search Work on Construction Documents? walks through RAG in more detail.
Why project records favour RAG
Project records have a few characteristics that matter here.
They change constantly. Drawings are revised, RFIs are answered and transmittals are issued every day. A fine-tuned model is frozen at the point it was trained. A RAG index can be updated as content arrives.
Answers need sources. In a contractual setting, an answer without a reference is of limited use. RAG returns the document, revision and passage behind each point so people can check it.
Permissions vary. Different people can see different projects and document classes. RAG can filter retrieved content by permission at query time. Knowledge baked into a model's weights cannot be withheld from particular users.
Content is sensitive. Fine-tuning creates a model that carries your data inside it, which raises questions about where that model lives and who can use it. RAG keeps the content in your store and the model separate.
Where fine-tuning can still help
Fine-tuning is useful for teaching a model a style or a task. Examples include classifying documents into your organisation's document types, extracting ISO 19650 fields from title blocks, or writing summaries in your house format. In each case the model learns a behaviour, and the facts still come from retrieval. For most organisations, a strong general model with good RAG covers these needs well enough that fine-tuning is unnecessary at the start.
What makes RAG work well on CDE content
RAG is only as good as the content it retrieves. Construction records bring particular challenges.
Scanned drawings and letters need OCR so their text can be indexed.
Metadata from different CDEs must be mapped to a common schema, so a revision or status means the same thing everywhere.
Superseded revisions need to be labelled, so the AI favours current documents unless asked about history.
Large documents should be split along sensible boundaries, such as sections and sheets, so retrieved passages make sense on their own.
Content from every platform should sit in one index, so questions can span the whole project.
Our article on preparing CDE data for AI covers these steps.
Where it should run
For most construction and infrastructure organisations, the index and the model should run inside the organisation's own cloud tenancy, in a known region. That keeps data residency, SOCI and contractual obligations straightforward and lets existing security policies apply. Content from Forma, Aconex, ProjectWise, Procore, SharePoint and others can be brought into that tenancy and kept current, giving RAG a complete and consistent record to work from. Teams that want to connect AI assistants to the index can use MCP, the open standard for exposing search tools to AI applications.
RAG and the Utopiaverse™
The Utopiaverse™ is built on this approach. CDE Sync™, CDE Backup™, CDE Migrate™ and CDE Archive™ bring content from every CDE into one connected record in your own Azure. AI Insights runs natively on Azure in your tenancy, using RAG and OCR search to answer questions across every project and platform with citations to the source document. The file viewer opens any cited document directly, and dashboards and reporting show document control status across the portfolio.
The organisation gets AI that stays current as projects move, answers people can verify, and visibility across every project in every CDE, with the data staying under its own control.
See RAG across CDEs in the monthly live demo, or talk to the team about your AI approach.