SkillsFuture Singapore (SSG) is experimenting with the Ada Embedding GPT model as a way to potentially improve the searchability of its Myskillsfuture website.
CEO Kok Yam Tan wrote in a LinkedIn post that the “rough prototype” search engine appeared to “respond better to creative search strings” and return a “much tighter” set of results.
He asked others to try out the experimental search function - which is hosted away from the production instance of the website.
Tan wrote that the agency wanted to “magick” the prototype with Singapore technology agency GovTech in order to “come up with a product good for deployment.”
He said that the early development work was completed by a mix of SSG and GovTech personnel.
Tan added that while the application is “just course search”, it is but one idea that SkillsFuture Singapore has for using large language models or LLMs.
LLMs have burst into the public consciousness since late last year when ChatGPT emerged. ChatGPT isn’t an LLM itself but acts as a simple way to query one.
“More broadly, we think there is great potential to use LLMs in providing upskilling advice, and maybe even in career guidance,” Tan wrote.
“This new tool to serve people better is exciting.”
Tan added that the agency’s experiment had also produced challenges, including “latency, how to incorporate structured data like course ratings [and] hallucinations.”
Hallucinations, in particular, have emerged as a major hurdle for enterprise and government organisations that are trying to make use of LLMs and ChatGPT-like interfaces.
The risk is that the AI confidently produces and displays a result that is completely fictitious.
In ChatGPT, the problem is particularly prevalent when seeking references - the source of results that the tool is returning.
Tan added that the agency is pleased to be “learning by doing”, adding that “that’s part of what skills-future means.”
