Meet the Data Science for the Public Good team for this project

  • Undergrad Interns: Sherlock Chen and Nebiyou Mengistu
  • Graduate Mentor: Xiaoyi Zhao
  • Faculty Advisors: Michael Cary, Yujuan Gao, Le Wang, and Steve Wang​

Project description 

As the adoption of chatbots has increased substantially in recent years, ensuring the accuracy and reliability of their responses has become increasingly important. However, hallucinations and limited transparency of data sources from conventional chatbots lead to inaccurate information and misleading responses. For Virginia Cooperative Extension (VCE) agents, hallucinations or misleading responses from chatbots could negatively impact the communities they serve. 


To address these limitations, we developed the VCE AI Tool, a chatbot that combats these challenges through the implementation of retrieval-augmented generation (RAG). Instead of relying on pretrained data, the RAG system restricts the chatbot from generating responses and sources based on the specific data it is provided, mitigating the risk of fabricated information as well as improving source transparency. We achieved this primarily through the R packages of Ellmer for calling large language models and DuckDB for querying specific data to answer prompts. To further enhance information retrieval, we created metadata JSON files to provide descriptions and interpretations of the data. While previous research has built chatbots to answer policy questions on a national level, our chatbot focuses on the more granular, county-level scale for Virginia cities and counties.

"By combining retrieval-augmented generation with trusted county-level Virginia data, the VCE AI Tool helps bridge the gap between available data and actionable knowledge. Rather than requiring Extension agents to spend significant time navigating complex data portals and analytical tools, it enables them to access locally relevant insights more quickly through an intuitive AI interface, while maintaining transparency by linking responses back to trusted data sources. This can strengthen the quality and timeliness of information agents provide to communities, expand the use of data in local decision-making, and improve VCE’s capacity to identify and respond to emerging needs across Virginia. More broadly, the project demonstrates a responsible approach to using AI within Cooperative Extension—augmenting agents’ expertise while making reliable, place-based information easier to access and apply.” - Le Wang, professor and David M. Kohl Chair
3 people stand in front of a research poster.
Two students stand in front of a screen presenting
Two students presenting

Internship insights

Sherlock: My time as an intern at Data Science for the Public Good (DSPG) has not only taught me technical skills, including programming in R and working with chatbot applications, but also how to build technology that directly meets the needs of the intended users. One particularly memorable moment was participating in the stakeholder workshop, where I learned the importance of asking open-ended questions and ensuring that both our team and the stakeholders shared the same understanding of the project’s goals and expectations. For this project, that meant designing a chatbot that VCE agents could easily understand and use to access complex data and use that information to respond to emerging local needs.  

During our first stakeholder meeting, my teammate and I applied these principles by asking the stakeholders questions beyond simply creating an operable chatbot. For example, we discussed how important response time was to them and gathered their preference on follow-up prompts for potentially ambiguous questions. Based on these conversations, I was able to integrate this big picture of the project into my workflow by emphasizing health, education, and agriculture datasets as well as designing an intuitive, user-friendly interface that would make the chatbot accessible to VCE agents. This internship has given me a deeper understanding of both the technical and human aspects of developing technology that is not only functional but also practical to its intended users. 
 
Nebiyou: During the Data Science for Public Good (DSPG) program, a standout experience for me was the Retrieval-Augmented Generation (RAG) workshop. Dr. Steve Wang frequently posed interactive scenarios, challenging us to anticipate potential issues rather than just feeding us answers. This engaging approach set the tone for me for the rest of the summer, encouraging me to embrace independent problem-solving and critical thinking. 

While I already had a foundational understanding of AI, the RAG workshop proved to be crucial when my teammate and I faced architecture issues after we built our metadata for our chatbot. We applied these insights to engineer a conversational agent in R. Using new packages and vector embeddings, and more importantly, learning how to use them, made me feel like I significantly advanced my AI engineering skills. This hands-on application felt like a natural progression from my previous work with Langchain, where I made another chatbot in Python. 

Beyond AI development, the library’s Git and GitHub workshops definitely enhanced my professional readiness. Previously, I only used Git for solitary commits on nearly finished projects. Learning to navigate collaborative version control – managing pull requests, resolving merge conflicts, and coordinating branches – equipped me with essential technical teamwork skills.  
As I prepare to graduate and either pursue a master's degree in data science or applied AI (or pursue full-time work in said fields), being able to bridge the gap between theoretical concepts and applied, collaborative engineering has made me feel increasingly prepared for the corporate workforce. 

Graduate mentor insights

Xiaoyi Zhao

One of the best moments that I have been through is the time that I spent with the interns when we needed to tackle issues together.

In the process of developing the AI agent for our program, we have been getting help from the faculty advisor along the way in improving the tools and workflow.

However, most of the time, the students and I need to sit down and untangle all the errors that we encountered along the way and look for better solutions. That is the point where I get to get a deeper understanding of the agents, the working mechanism, and the underlying logic of open models. That is also the point at which the students start to understand why we need to use the Metajson for sustainability, why we use the tools instead of letting the agent write their own answer, how every line of the prompt to the model contributes to the final answer, and the motivation for why we want to set the goal for this agent to incorporate different formats of data and using vector embedding. The actual growth of the team happens when solving issues and brainstorming to improve the agents.

What our stakeholders had to say about their experience

“The students working on the Data Science for the Public Good project did a fantastic job listening to the needs of the for Virginia Cooperative Extension client to develop an easily accessible and trustworthy AI tool to compile data about the communities in which agents and specialists work. They took time to ask meaningful questions about our needs and incorporated them into the project scope, which eliminated need for re-work or creation of an unusable product. The final version was demonstrated to the users, and it had a clean look which included exactly the data sets that were to be used which added a huge amount of reliability to the output. I look forward to using the product and am excited about the commitment of ongoing work to refine and expand the scope.”

Scot Ferguson, Extension Agent, Agriculture and Natural Resources

Click here to view all of the 2026 DSPG projects. >>