The Tilburg.ai chatbot is currently being rolled out to more and more courses at Tilburg University and will soon be rolled out to the Eindhoven University of Technology. Have you already used it yourself? Fantastic! Then you have probably already noticed that our chatbot is different from what you are used to, for example from the chatbot experience of ChatGPT. Although our chatbot uses the same underlying AI model, we have added several functions that improve your learning experience. How does this work exactly? And above all: How does this help you prepare for your exam? We will get you up to speed in this article.
Directly Connected to Course Material
The chatbot is directly connected to your course material. Your professor has added the course content to the chatbot by uploading the files that you, as a student, normally access on Canvas, making a dataset available behind the scenes. This allows you to ask questions about the entire course material uploaded. When you pose a question to the chatbot, it locates the sources of the answers, similar to how you would browse through the course material on Canvas to find the right information. Below, you can see an example of how this process works.
Let’s say you ask the following question:
"What is the methodological framework from the Fields Of Gold paper?"
The chatbot splits this main question into multiple sub-questions. By splitting the main question into multiple sub-questions, the chatbot can search more effectively for relevant information in the documentation. It generates targeted sub-questions, with these questions the chatbot can explore multiple facets of your question rather than relying solely on the original question. For each sub-question, the chatbot searches through the course materials.
This multi-question search strategy increases the likelihood of retrieving the relevant information, especially for questions that require more steps to answer or involve ambiguous terms. Additionally, by dissecting the main question into these specific sub-questions, the chatbot is already thinking about how to answer your question and what approaches to consider, meaning that each aspect of the answer (each subquestion) is addressed.
For our example, the following sub-questions are generated:
- What is the purpose of the methodological framework used in the Fields Of Gold paper
- How does the methodological framework support the research objectives?
- What are the key components of the methodological framework in the Fields Of Gold paper?
Great, now the sub-questions have been formulated, the chatbot will look to identify where the answers can be found within the course material. While there may be a document with the field of the gold paper, your professor might have also made some remarks about it in his or her slides. Therefore, the answer to your question often consists of multiple documents and can be spread across several sources. It is also possible that:
- The same information is mentioned in multiple places within a document.
- Answers are similar but are formulated slightly differently.
- Multiple sections can be relevant to a single sub-question.
In this example, however, we know that the information is contained in a specific paper: FieldsOfGold.pdf (the title of the source already hints at its content). Therefore, the filters applied are related to the source.
- source: ‘FieldsOfGold.pdf’
- authors:’Johannes Boegershausen, Hannes Datta, Abhishek Borah and Andrew T. Stephen’
For our example here, we are discussing a case where the answers may be scattered across the source. The chatbot will look for the most relevant parts of the text. The system does this by assigning a relevance score to different sections of the text. The chatbot then selects, for example, the top 10, most relevant parts of the text and formulates an answer based on the information written in these parts. This process ensures that the chatbot provides more accurate answers to the questions posed to it.
How to Use Our Course Chatbot
This procedure explained above is called RAG (Retrieval-Augmented Generation). Thanks to this process, new avenues for asking questions have opened up. In contrast to the traditional way of interacting with ChatGPT, with RAG you can formulate your question much more specifically and targeted. Let us explore these different possibilities one by one.
Direct Explanation of Each Slide
If you get stuck while learning or preparing a class, tell the chatbot where you are stuck. Instead of asking ChatGPT a question about the overarching subject, ask the chatbot very specifically: “Explain figure 5 in paper X,” or “Explain topic Z in relation to the rest of the slide in slide Y of week O.”
The chatbot works similarly to a search engine. Instead of asking for explanations about a general topic, you can tell the chatbot very specifically and precisely where in your course material you want added information about. With the RAG system, it not only retrieves the relevant information but also understands the context and therefore gives a more accurate answer.
Example
During the second week of the Organization and Strategy course, you come across a rather complex figure that is only briefly explained in the course material. But without any additional explanation, you can’t understand the information presented or you need additional examples that clarify the underlying concepts. Instead of asking ChatGPT a general question, such as “What does this figure mean?” with a picture of it, you can pass the name of the slide and its title to the chatbot. That’s all you have to do: the chatbot will look up the relevant figure, investigate the context, and give an answer that understands the context. This way, you will receive the most appropriate answer based on the course material, as illustrated in the accompanying video.
Get Assistance with Practice Questions
We can extend the previous example to practice assignments. Suppose your teacher has just uploaded a series of practice assignments and you get stuck on a specific question. You can use the chatbot to get started step by step: first name the question and explain which part of the assignment is unclear. Then ask the chatbot to retrieve the relevant parts from the course material to answer the question.
Do you need additional information, such as a step-by-step plan, a similar assignment, or extra examples? Then clearly indicate this with your question. The chatbot will search for the necessary sources and substantiate its explanation with references or fragments from the study material.
In this way, you work together with the chatbot to work on the assignments in a didactic and effective way, without having to spend hours yourself going through all the documents. However, make sure that you actively work on the practice assignments yourself; simply getting the answer is not the goal!
Source of Information
You might wonder, “What if I don’t know exactly where the topic is in the slides?” Or: “What if I just want the main topic explained?”
To accommodate these reverse cases and improve the transparency of the answers, we have built in another feature. At the end of each answer, the chatbot always adds the source(s) on which the answer is based. This means that you can see exactly where the information comes from, such as specific slides, figures, or pages from a paper.
Why Do We Think This Is Important?
Transparency improves the trustworthiness of a chatbot. Large language models, such as ChatGPT, answer questions based on probability (read: they are not factual based). It is therefore important to verify the information provided by the model. With our chatbot, you always know what the AI model bases its response on. This makes the Tilburg.ai chatbot answers verifiable.
In addition, the ability to view the sources means that if you want to learn more about the topic or understand the context better, you can easily go back to the source and read the details yourself. Finally, going back to where we started, even if you don’t know exactly where the information is,the chatbot will direct you to the right place in your course material. This saves both time and effort.
Summary
In short, the Tilburg.ai chatbot transforms the way you interact with your course material. Where PowerPoints, papers, and other documents used to be static, you can now work interactively with your course material. The RAG method ensures that the chatbot not only retrieves relevant information but also places it in the right context, resulting in accurate answers.
Moreover, the chatbot is transparent: by always mentioning the source of the information, you maintain control over your learning process. You can verify the answers and, if necessary, examine the material in greater depth. We will continue to develop the chatbot to improve your learning experience with the chatbot. So, in the near future, follow us on social media, and we can’t wait to receive your feedback!
