©JackF – stock.adobe.com
OpenAI’s Chat GPT product has captured the imagination of the public, but finding practical use cases in construction is challenging.
Unlike AI used for decision support or automation, the technology is not explainable—there is no audit trail to pin down how the application decides what to generate, why or where information is sourced from. This makes it difficult to use in some situations where events leading up to a decision must be traceable for liability or other purposes.
But there are still viable routes to market for the technology. Preconstruction artificial intelligence (AI) application Togal.AI now harnesses ChatGPT—but not to generate unique text outside of a chat environment or to replace its AI takeoff functionality—but as a natural language search tool for construction documents. This keeps the AI focused on finding and presenting data from authoritative data in design drawings while improving the productivity of the preconstruction team.
According to Togal.AI President and Co-Founder Patrick Hughes, this is not only a nice value-added feature but will build on the company’s existing momentum as the company grows its customer base for AI-driven takeoff tool.
“Our number one priority is continuing to build an intuitive takeoff software that anybody can log on and understand in five to ten minutes,” Hughes says in an IRONPROS Product Deep Dive on Togal.AI. “We combine this with a lot of deep AI and very sophisticated algorithms that help enhance the end user’s efficiency. We are doing some really complicated stuff. It is important to make that easy and consumable for the user … We took ChatGPT and are making that easy and adoptable in construction documents. If we do that, then we continue to grow our customers, get new customers, expand in the U.S. and internationally and sign partnerships.”
Semantic Search for Construction
By harnessing ChatGPT, Togal.AI provides semantic search functionality for construction document management. This means documents are searched based not just for specific search strings in text or images, but by examining the intent of the material in relation to the search query.
“The big picture is that generative AI releases like ChatGPT or Google Bard are transformational to our economy and to all industries,” Hughes said. “Construction has been slow to adopt technology for a variety of reasons. There is a higher ceiling and more room to grow in construction than any other industry. But one barrier to growth is the sheer amount of data and how little of it is carried on to the next job. The process and power and not been affordable or readily available to take in and digest all of this. Whether it is warranty issues or budgets or schedules to weather conditions or personal issues, contracts with subs, contracts with owners, with ChatGPT, we believe we will be able to combine all of that and provide a usable output.”
While ChatGPT is not an explainable construction AI technology—there is no way to audit or reverse-engineer what it does—search within an application structured and unstructured data domains is a low-risk use case for construction.
“In our first release, we are only giving the capability for the GPT to search what is uploaded—not accessing the entire internet or every plan known to mankind,” Hughes said. “But this is a Bayesian search of a more refined set of documents than the entire internet—applied to a smaller subset of materials and it is easier to sort that information.”
Togal.AI Head of Operations and Investor Relations Karlie LaCroix described in correspondence with IRONPROS the two distinct processes ChatGPT can now perform in the product and why then opted for ChatGPT versus generative AI products available from other vendors.
“We have picked two models,” LaCroix said. “One for ‘ingesting’ the data, also known as an embedding model. We are also using a generative model for text generation. We can mix and match them, but we decided on both models from OpenAI. The honest answer is that we went with a leader in the field, but this is also backed by the numbers.”
Here, she points to work done by Predictionguard Founder Daniel Whitenack comparing how accurate 27 large language models were in spotting semantic similarities.
Learn More about Togal.AI on IRONPROS