Keywords
chatbot, biosampling, rag

Technologies
node, expressjs, transformers, smollm

This project examines the merit of using a generative AI chatbot to onboard participants for a biosampling research study. Recruitment and retention for biosampling studies is affected by a variety of psychosocial factors, including participant embarrassment and distrust. Chatbots can reduce feelings of exposure and could thus improve participation rates. A chatbot prototype was developed to guide participants through a stool sampling procedure and answer questions. The chatbot backend uses Node.js and the Hugging Face transformers.js library to run a local instance of SmolLM2, a lightweight text-to-text generative model. Prompt engineering and retrieval-augmented generation are used to improve response relevancy, with content gathered from authoritative healthcare sources. 60 participants in a comparative study were assigned to onboard with either a human researcher, or the chatbot. The chatbot was found to reduce embarrassment and maintain comprehension, but participant experience was negatively impacted by technical limitations such as slow response speed and repetitious dialogue.