Hi everyone
I’m working on a construction estimating workflow where we tried integrating ChatGPT into our material takeoff and cost estimation tool. The idea was to make the process easier for contractors who are not technical and want quick breakdowns of materials from drawings.
However, we noticed that when the inputs get a bit complex (like multiple layers, different measurement units, or structural specifications), the AI sometimes gives inconsistent or simplified responses. We were expecting the model to maintain accuracy with construction terminology and quantity logic, but it doesn’t always pick the correct context.
We also tried adding custom instructions and examples, which helped to some extent, but still not fully reliable for real-world use.
So my questions to the community are:
-
Has anyone worked on domain-specific fine-tuning for construction or architecture projects?
-
Would embedding-based context retrieval improve accuracy in this kind of workflow?
-
Is there a recommended approach to maintain measurement consistency in generated outputs?
If it helps to understand our use case, here is the page that explains the assistant concept we are experimenting with:takeoffsharks.com
I’d really appreciate suggestions from anyone who has experience aligning GPT models with technical or industry-specific tasks
Thanks in advance!