TerpAI is a GenAI chatbot similar to popular GenAI chatbots you may have heard about. TerpAI is being presented to the UMD campus by the AI Solutions team.
This is part of an ongoing effort to encourage greater AI awareness and provide AI resources to the UMD community. Having a strong knowledge of AI tools such as TerpAI can help differentiate you from others in the job market, especially as GenAI's impact continues to grow.
TerpAI can be accessed through its online URL located at terpai.umd.edu.
The best place to start learning about TerpAI is at the AI-Resources@UMD website.
Within the TerpAI interface, you can select the on the bottom right side of the screen (next to the chat input) for a guided tour of the interface.
The AI Solutions team is part of the Division of Information Technology. Our team supports TerpAI among other AI products and services.
The AI Solutions team does not encompass all AI work being done on campus. The campus has many diverse groups that offer services, do research, and educate the UMD community in AI. If your request lies outside the team's purview, we can direct you to other on-campus resources.
You can always email the AI Solutions team at dit-ais@umd.edu.
In a nutshell, GenAI chatbots can comprehend user questions and generate human-like responses.
Although GenAI chatbots are the most widely used type of GenAI tool, there are other GenAI tools which generate images, speech, music, and more!
It is quite difficult to explain how GenAI works without getting technical so here is some background information:
Non-deterministic means you may get somewhat different answers even if you ask the exact same question to TerpAI twice.
GenAI models, such as those used in natural language processing and image generation, are often non-deterministic. This is considered a feature.
Here are a few reasons why this occurs:
Stochastic Processes: Many generative models incorporate randomness as part of their design. For example, during the generation process, models might sample from a probability distribution to make decisions about which word or feature to generate next. This introduces variability in the output.
Training Variability: The process of training generative models often involves stochastic gradient descent or similar optimization techniques, which can lead to different model parameters even if the training process is repeated with the same data.
Temperature Settings: In models like GPT, the "temperature" parameter can be adjusted to control the randomness of the output. A higher temperature results in more random outputs, while a lower temperature makes the output more deterministic. Users can adjust this setting to balance creativity and coherence.
Beam Search and Sampling Methods: In tasks like text generation, methods such as beam search or sampling with top-k or nucleus sampling introduce elements of randomness to explore different possible continuations of a sequence.
These factors contribute to the non-deterministic behavior of generative AI, allowing for creative and diverse outputs, which can be particularly useful in applications requiring variability and innovation, such as creative writing, art generation, and more.
Other popular GenAI chatbots have stricter limitations on their free versions. You won’t get unlimited use of TerpAI but it has a larger quota than many free GenAI chatbots.
A prompt is the input or question sent to a GenAI chatbot. Here are some sample prompts
“What is Generative AI?”
“Write a rap about the pressure students feel during final exams in the style of Eminem.”
In TerpAI, you can type in your prompt at the input once you have initiated a chat with an Agent. Once you hit Enter, TerpAI will provide a response to your prompt.
Prompting is indeed an engineering, one that involves the practice of crafting clear and concise inputs or prompts to elicit more reliable output from GenAI Chatbots like TerpAI. TerpAI possesses a Quick Prompts feature that allows you to retain a collection of frequently used prompting language and build your own prompt library. Additionally, you can find a collection of prompts organized by categories at AI Prompt Library
A search engine query primarily responds with a list of URLs ranked in order of relevance.
By contrast, TerpAI gives human-like responses to user prompts. It's like talking with a very knowledgeable person.
Unlike search engines, you can ask followup questions to TerpAI. TerpAI retains the context of earlier conversations within the same chat session.
For example, you could prompt TerpAI with "What is a black hole?" and get a response. If you then prompt "Explain like I'm five", TerpAI will give a simpler explanation. It recalls that you asked about black holes.
Or you could follow up with "Is it the same as dark matter?". TerpAI understands that "it" refers to a black hole.
With TerpAI, you can have back and forth conversations. TerpAI recalls past conversation within the same chat session. When you start a new chat session, TerpAI begins with a clean slate. It will not have knowledge of previous chat sessions.
Here's one thing search engines can't do: suggest ideas. Let's say you're asked to give a speech on AI use in higher education--TerpAI can help you outline that speech. Or maybe you need ideas for dinner this week. Ask TerpAI to recommend recipes based on the ingredients you have at home. If you don't like the recipe, ask TerpAI to make changes!
You might expect TerpAI to give a perfect response. It won't and you shouldn't expect it to. It won't write a perfect speech the first time. What TerpAI will do, is get you started on that speech. Sometimes, all we need is a little push to accomplish our goals!
TerpAI can do all this and more!
For the most part, yes. However, GenAI chatbots, like TerpAI, can hallucinate.
Hallucination in a GenAI chatbot response occurs when it provides answers that range from somewhat inaccurate to totally fictional. Hallucination isn’t specific to TerpAI but happens to all popular GenAI chatbots. You should always double-check the response especially if it sounds suspicious. Unfortunately, this is an outcome of the neural network training. Fortunately, hallucinations are infrequent and newer GenAI models aim to reduce hallucinations.
TerpAI does not have the most recent information. TerpAI uses GPT 5.2, the same model that ChatGPT uses. The last time this model was updated was September 2025. The model could get updated at some point, or a new model may come out. TerpAI may have real-time information soon. Keep an eye out!
TerpAI has general knowledge of many things, but it has not been specifically trained on UMD websites. TerpAI uses the GPT 5.2 model which has vast amounts of general information, some of which includes UMD. However, because it has general information, TerpAI can miss UMD-specific information like lectures given on campus, announced snow days and other frequently updated information.
You should confirm with your instructor about their policy using TerpAI (or any GenAI) for your courses. Also be sure to read through the acceptable usage guidelines when using TerpAI.
Yes. The one exception is emeritus professors who continue to have access. Losing access to TerpAI also means losing access to a your chat history.
No, administrators do not have access to chat histories.
TerpAI is approved for up to Level 2, Moderate Risk data, based on the University of Maryland Data Classification Standard.
Any data you put into TerpAI will not be shared with other users or used to train the AI. While session data will remain in UMD's Azure tenant, questions that utilize the internet search feature will go through a 3rd party vendor's Azure tenant.
It means your access to TerpAI is not unlimited.
Why? Due to the immense computational requirements of GenAI chatbots and the cost to run these tools, most free GenAI chatbots place limits on their use, including TerpAI. You are given a daily quota to limit that usage. Each interaction with TerpAI decreases the quota you have. If you ask TerpAI too many questions or make too many requests, you will reach your quota. When that happens, you won't be able to use TerpAI. You'll have to wait until your daily quota is reset at 7 pm EST.
For students, there isn't a way to do this now.
For faculty and staff, read further down in the FAQ on how to get additional quota below, starting at: How do I switch to another TerpAI Agent?
Contact the AI Solutions team to discuss other ways to increase the quota by emailing: dit-ais@umd.edu.
It takes a substantial amount of technology and computational resources to run TerpAI. This is true for all widely used GenAIs. To avoid using too many resources, GenAIs, like TerpAI, limit how much you can interact with them through Daily Quotas.
By contrast, search engines require less computation and therefore less money to run. Google made its search engine freely available by selling ads to customers that wanted to target their ads to certain relevant queries. Other search engines have followed a similar approach. GenAI is far more computationally expensive compared to search engines.
Go to TerpAI@AI-Resources where you'll find a chart that shows the maximum number of interactions of types of file you can use in one day. Use the chart from that to determine how many words you can use in a day. As that value might change in the future, we'll say the quota is N words where N is the number in the chart.
Quotas are allocated on a per-agent basis, with usage measured against each active agent associated with supported models, currently one of GPT-5.2, GPT-4o, and GPT-4o-mini.
No, it's not free to upload images just because we use the term "words".
Think of "words" as the unit of measurement for the daily quota. Those who are more familiar with GenAI technology have heard the term tokens. Think of going to an amusement park where you pay some money, and get some tokens. These tokens are the "currency" to get on rides or play games. There's a similar concept in gambling where you exchange money for chips, and play games of chance with chips. Those are basically tokens.
GenAI tokens are similar. Roughly speaking, one token is about one word. The precise definition of a token is technical and not that important. We use words to make it easier to understand, but think of these words like a unit of currency.
Image uploads reduce your daily quota by a variable amount (up to 400 words), capped at TerpAI’s maximum supported image size.
A typical interaction involves a user prompting TerpAI and TerpAI giving a response. Both the prompt and the response decrease the quota. In general, prompting TerpAI uses far less quota than TerpAI's response. As noted earlier with images, file uploads also come with a cost. The costs are much higher for a non-image than an image. The limits by file type can be found in a chart located at TerpAI@AI-Resources
N words represents the total quota. If you have ask TerpAI questions directly through a prompt or chat with a file and then ask TerpAI questions, all of this is part of the quota. For example, if you use 20,000 words using text interactions, that leaves fewer words for loading a PDF.
It is possible to consume you entire daily quota in one prompt if the chat history is large enough! The answer to why is fairly technical. TerpAI can only retain the information in the chat to continue conversations for a limited time. When you resume a chat the following day, TerpAI has to reload the conversation from the prior day so it can continue the chat session.
The reload isn't free. It reduces your quota. You can avoid this reload charge if you start with a new chat. We highly recommend starting with new chats each day unless you really need to continue an old chat session. You should start new sessions if you are away from your computer for a while to avoid this problem.
If your daily quota is used up, you will be informed that you have to wait until the next day to get the equivalent of N words. This happens at 7pm EST.
Quotas do not carry over and reset back to the full quota at 7 pm EST each day. Unused quota from prior days expire.
Go to TerpAI@AI-Resources to find a list of TerpAI features. Here are a few other features.
It can understand prompts in many languages, including French, Japanese, Tamil, and Filipino. It responds in the language of the prompt.
It can translate text from one language to another. You can prompt it by saying “Translate the following to English: 日本の首都はどこですか. (It translates to “What is the capital of Japan?”).
Not only can TerpAI analyze code, but it can also generate programs in the most popular programming languages. Make sure to follow rules for a class that might restrict the use of TerpAI or any GenAI from being used in specific situations.
You can think of an idea for a song, then have it write lyrics in the style of a singer. Or ask for the song to be written like Dr. Seuss.
TerpAI recalls its answers from the same chat session so if you need clarification, you can provide follow-up prompts and interact back and forth.
TerpAI understands many languages besides English. It can understand French, Japanese, Tamil, and Filipino among many, many other languages. TerpAI will respond to prompts in the language of the prompt. For example, if you prompt TerpAI with Japanese characters, TerpAI will respond with Japanese characters.
Another feature of TerpAI is to translate text from one language to another. If you don't know Tamil, you can ask TerpAI to translate Tamil text to English.
One strength of GenAI chatbots is its ability to analyze and generate programs based on your prompts. By crafting the right prompts, it is possible to develop deeper and broader technical aptitude by allowing TerpAI to generate exercises, provide step-by-step guidance and assist in debugging and breakdown of existing code. TerpAI's breadth of knowledge is extraordinary in this regard.
Yes. To use the voice feature in TerpAI
The speech-to-text only handles English. It lacks support for other languages. However, you can type prompts in other languages.
Once you get a response, at the bottom, there is a set of icons.
Click on to have TerpAI speak its response.
This is one feature where TerpAI can't handle multiple languages. It can speak English. However, TerpAI pronounces French answers as if they were English words which means it is being mispronounced. For languages that don't use the Latin alphabet (English and French use the Latin alphabet) such as Tamil and Chinese, the speaking feature does not work at all.
To chat with an image, you will need to upload the image and submit as part of your question. To accomplish this:
This is a feature of TerpAI that allows you to upload a single file, including images (JPG or PNG), Word documents, Excel spreadsheets, and PDFs.
The current supported image file formats are jpg and png.
Note that you can also drag and drop an image to populate in input. Once uploaded, you'll see the file name, in this case, motorcycle.jpg.
You can ask TerpAI a question about the file or use the file for a specified task.
You can ask followup questions. TerpAI is able to determine the type of engine and build for the motorcycle. TerpAI's ability to analyze images is quite impressive!
Similar to chatting with an image, you can upload any document file type (Word, PDF, Excel) by clicking on the paper clip icon. Upload any of these files, then add a follow-up prompt.
TerpAI utilizes a Bing Search API that enables web search capabilities. This add-on allows results to be enhanced with the latest news, media and releases beyond cutoff knowledge information. Currently, it is available by default on primary UMD agents, but it is possible to create a locked down personal agent and define a more controlled internet experience.
GPT stands for Generative Pre-Trained Transformer. It’s basically a neural network that’s been previously trained on large amounts of text. That's the pre-trained part.
From this information, a GPT system generates a response. That's the generative part.
Finally, transformer refers to the kind of neural network used to understand language when you enter a prompt.
To view available agents:
For students, they will see GPT 5.2 - UMD. The dropdown only contains this one choice for students.
Faculty and staff can pick from one of two GPT systems. The default GPT system is GPT 5.2; the other is GPT 4o-mini.
Here are the steps to use GPT 4o-mini assuming GPT 5.2 is the current selection.
After you accept the guidelines and enter a prompt, the chat session is set to the selected GPT system. For example, if you select GPT 4o and begin chatting, the dropdown disappears.
You can't switch to 4o-mini in the middle of a conversation. You need to start a new chat. Furthermore, you can't continue the old conversation that was originally in GPT 4o in GPT 4o-mini.
You'd have to ask TerpAI the same questions again in GPT 4o-mini which would decrease the GPT 4o-mini quota.
GPT 4o delivers top-tier performance, handling complex tasks and providing highly detailed responses. GPT 4o-mini offers a faster, more efficient experience, perfect for quick answers and lighter tasks, while still maintaining great quality. Both are designed to suit different needs seamlessly.
GPT 4o has a daily quota, and GPT 4o-mini has a separate daily quota. Each chat session in your chat history is either associated with GPT 4o or GPT 4o-mini, so previous conversations retain the GPT system and do not change when you change systems in a new chat.
if you're faculty/staff, and you run out of quota on GPT 4o, you can switch to GPT 4o-mini. Once you've used up both quotas, you have to wait until 7 PM EST to have the quotas reset back to baseline.
When you enter a prompt, you may notice a icon
This appears on the right side of the chat input bar at the bottom of the TerpAI screen. It appears where the button would normally appear.
This spinning icon indicates an in progress status while TerpAI is working on a response. If you click on the spinning icon, TerpAI will stop working on the response. This is a good idea if the spinning lasts a long time. Once you stop TerpAI's response, you can start a new chat, and the status should reset.
When you navigate away from an active TerpAI tab while it is generating a response, the system will pause its output. It will resume only when you return to the tab. Please be aware that this may lead to issues with TerpAI's configured timeouts.
AI technology is quickly evolving.
TerpAI is expected to make future improvements. This FAQ may have to be changed from time to time to reflect those updates. Look out for future enhancements!