RAGs to Research: Exploring Applications for Maizey in Undergraduate Research

Joey is an intern with ITS Marketing and Communications. One of his projects this summer is to develop and test a unique Maizey. Joey chose to bridge the gap between seasoned researchers and overwhelmed college students by creating a research assistant chat bot.

A mind-numbing pain that I share with many other students is one that, satirically, our University takes immense pride in. We are a top-ranked research school that funnels $1.86 billion annually into research. The Undergraduate Research Opportunity Program (UROP) is seen as a must-have for resumes, but so many of us are just tossed into a lab and expected to know how to research. Professors lighten the load for us by having us start with lighter tasks like literature search, but what happens when reading research papers is that you “read but forgot to understand.” Having the opportunity to create my own Maizey project — a straightforward process that any student can do in minutes — I decided to create something that would help bridge the gap between seasoned researchers and overwhelmed college students. A chatbot for researchers.

An elderly anime man buckled into a car seat looking shocked. Text reads: When you read a whole page and then realize you forgot to understand.
Meme created by Joey Lin on imgflip.com
on a beige background, hand extending from top right of the frame pinching a white dangling terry cloth
Adobe Firefly generated image

Maizey is a RAG. Not that we wipe tables with, RAG stands for Retrieval-Augmented Generation, and is a distinct type of generative AI model. Maizey takes input from a Google Drive folder or Canvas course, and can read most text-based file types like .pdf or .docx. So when you ask it a question, Maizey will go to the input source and look for relevant information to append to the query. It then runs that augmented query through a large language model like GPT 4 Turbo, and returns the response to you. This makes Maizey capable of some pretty terrific things, as it really is what you make of it.

In my attempt to replicate greatness, I fed my Maizey the papers we read for a lab I was a part of and started asking it questions that our lab manager often asked of us. I was elated that it was responding until I asked U-M GPT the same questions. It gave the same, if not a better, answer, meaning that my Maizey did not learn anything from the papers I fed it, which is understandable. I do that, too.

A side-by-side of Maizey's response and U-M GPT's response to Joey's question about cultural tightness. The responses are nearly identical despite Maizey having direct access to the papers from Joey's lab.
Maizey (left) and U-M GPT in a battle of wits.

But I wanted it to be better than me. I then tried to play around with the system prompt, which turned out to be quite the game-changer. I essentially told Maizey to roleplay as a knowledgeable research assistant, explained its capabilities in summarizing literature, explaining methods, identifying variables, etc., and told it to actually read the papers I provide. Then, it started giving pretty specific answers that it would not have known without reading the papers. One great feature about Maizey that works hand-in-hand with research is how it cites the material it retrieved additional information from, so it’s easy to track citations.

But remember how Maizey is text-based? With its current capabilities, Maizey does not deal with graphs, tables, or images that well. While Maizey can breeze through a text-heavy paper, it struggles to correctly return information from graphs. This means that at the time of writing, the efficacy of your research assistant Maizey, depends on the research papers in your field. This is something I will keep working on, as graphs and tables are crucial to papers and can be almost as confusing as text at times. Not to mention the countless times that it would just pretend it never read the papers and say, “I don’t know.”

Screenshot of Joey's Maizey responding "I don't know" to the prompt "what's the united states' cultural tightness looseness index according to the literature"
I appreciate the honesty, Maizey.

Throughout this experience, I’m starting to see Maizey’s potential as a research tool for both students and faculty. It reduces the time and resources needed for training and adaptation for students conducting research, such as those participating in UROP for the first time. It is useful outside of the lab as well, as Maizey is beneficial for research tasks that are essential in writing essays and completing other assignments. And although there are still obstacles to be surmounted and better prompts to be developed, Maizey’s use cases are growing day by day, and I’m excited to see it elevate the overall research and learning experience at the University.

My system prompt:

You are a highly knowledgeable AI designed to assist researchers by analyzing and interpreting research papers across various fields of study. Your capabilities include summarizing papers, comparing study results, explaining methodologies, and highlighting key findings and implications. When presented with a research paper, you meticulously review the content and deliver concise, accurate, and relevant information to address the researcher's questions. Your aim is to facilitate the researchers' understanding of the literature, support their ongoing work, and foster advancements in their field of inquiry.

Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.


Question: {question}

Helpful Answer:
Author: Joey Lin, ITS Marketing and Communications Intern

Joey is an intern with ITS Marketing and Communications. One of his projects this summer is to develop and test a unique Maizey. Joey chose to bridge the gap between seasoned researchers and overwhelmed college students by creating a research assistant chat bot.