The key to a chatbot’s usefulness is the data it is trained on. A professor at Dartmouth’s business school demonstrates that by creating a chatbot TA to answer student questions.
First, Shumsky created a knowledge base in Cody—an AI coding assistant, or online chatbot service—by loading the information from his course onto the website (cody.ai). The information included course readings, lecture slides, and transcripts of videos. Course participants could then prompt Robota for information by typing in questions, and Cody would call on ChatGPT 4.0 to respond with an answer based on the course materials Shumsky had provided.
humsky explains that the basic process behind the bot is that Robota will read the prompt question and then transform it into a group of numbers that represents a location in a “linguistic space.” Robota then searches the provided knowledge base for related text—that is, passages in a similar location in the linguistic space. Robota (via cody.ai) then combines those coded texts with the initial prompt and sends it to ChatGPT—the “engine” that converts the question and then provides an answer in English—and replies to the user.
(From this story by the Tuck School of Business)
He’s doing this not just to get help answer student questions but to practice working with large language models and other AI stuff, and to let students have experience with them: “The number of GenAI health-care applications being rolled out right now is overwhelming, and so I wanted them to have some interaction with these tools in this context so that they could start to understand what it involves.”
Return to the Concord Monitor