An Explication and Classroom Field Study of the Virtual Human Interaction Lab’s Expert (VHIL-E) LLM
Abstract
VHIL-E is the Virtual Human Interaction Lab’s Expert, a large language model (LLM) representing the lab’s research, teaching, and outreach on virtual and augmented reality. We first motivate the project and then present best practices for implementing a retrieval-augmented generation (RAG), specifically the “Seven Cs”: collecting, cleaning, classifying, chunking, creating embeddings, correlating embeddings into an index, and connecting the index to an LLM. We collected academic publications, transcribed public talks, dedicated interviews and news articles about the lab’s research, and various curriculum materials from Virtual People, the lecture course taught about the lab’s research for over two decades, resulting in over 2.3 million words broken into approximately 10,000 chunks. In Study 1, we compared performance on a multiple choice test of 231 questions between various implementations of the RAG system and traditional Base GPT models. Models generally performed well, between 83 and 90 percent, similar to student performance on course examinations. In Study 2, an open-ended task implemented over 10 weeks in the Fall 2025 Virtual People course, students (N=89) used VHIL-E to query and understand the course materials, and logged hallucinations which were “egregiously wrong answers.” Students then chose the single worst wrong answer over 8 weeks. They then compared the Base/RAG Hybrid—which prioritized the RAG but allowed ChatGPT to consult its general intelligence—to the RAG Constrained, which limited substantive information to the embedded index. Allowing VHIL-E access to GPT produced more than twice as many hallucinations as constraining it to the index. We discuss implications for scholars who build and use RAG-based LLM applications.