Frontier generative artificial intelligence models are advancing rapidly, but they still cannot accurately read the documents and drawings used in architectural design, procurement and construction. Searchdog, a startup, is solving that problem by adding search technology to large language models. Less than two years after its founding, the company has signed up major South Korean conglomerates as clients and now plans to expand quickly overseas.
Baek Jun-sun, chief executive of Searchdog, said in a recent interview with the Seoul Economic Daily at the company's office in Seongnam, Gyeonggi Province, that "graduate students and companies alike use retrieval-augmented generation and vector search technology, but no company has solved the problem of getting AI to read construction documents." The CEO added, "Searchdog started with documents such as contracts and requests for proposals, and now handles everything with AI, including two- and three-dimensional drawings and building information modeling."
Demand is strong in the design, procurement and construction industries for using AI to analyze such documents, but applying AI to actual work is not easy. Industrial engineering companies such as Autodesk, Bentley Systems and Aveva each use different file formats for drawings, and AI struggles to understand the volume of information contained in documents and drawings the way a person can. The construction industry's practice of revising contracts, specifications and official letters repeatedly, and the disconnect between documents and drawings, are also cited as obstacles. Baek said, "You can convert drawings into a common file format so existing AI models can read them, but about half of the information in the original is lost in the conversion. Even if users upload documents to existing AI models, they cannot get answers about the important information."
Searchdog resolved this by combining AI with various search technologies. Baek said, "Global industrial engineering companies are trying to solve this by tuning LLMs, but because this is a domain-specific field, search technology has to be applied. Searchdog extracts data such as shape, layer and dimension information from documents and drawings, then structures and connects it using search technology."
In practice, a user can upload a drawing file to Searchdog's solution and ask it to list the types and quantities of sanitary fixtures in a building, and immediately receive the number of toilets, sinks and shower stalls. That is possible even though the drawing does not label these items as sanitary fixtures, because the system analyzes the attributes of the information and its relationship to the drawing.
Baek said, "If contractors and design firms use this service, they can cut the time spent analyzing whether drawings comply with building codes and client specifications by 70%," adding that major domestic construction companies have been reaching out to propose proof-of-concept projects.
The service currently focuses on construction engineering but can be extended to other fields. The company has signed a supply contract with a large South Korean construction firm and is also working with a domestic semiconductor company. Baek said, "We were recently selected for Inception, Nvidia's startup development program," and stressed, "Our goal is to enter the U.S. market in the second half of the year."
