AI Resources & Training

1. Be specific with your questions: TritonGPT is designed to provide detailed and accurate responses, so try to be as specific as possible with your questions. Providing specific context will help TritonGPT understand what you're looking for and give you a better response.
2. Be conversational/ask additional questions: TritonGPT recognizes the back-and-forth nature of conversation. Don’t stop with the first prompt–continue to ask follow-up questions and ask for clarification if you don't get the response you were expecting. Your questions help refine its responses.
3. Use natural language: TritonGPT is designed to understand natural language, so try to ask your questions in the same way you would ask a human. Avoid using overly formal or technical language, and feel free to use contractions and colloquialisms.
4. Check the source material: TritonGPT has access to a wide range of documents and resources, but it's always a good idea to check the source documents linked at the bottom of the response. This can help you find the information you need more quickly and efficiently.

Use TritonGPT for a variety of tasks: TritonGPT is a versatile tool that can help with a wide range of tasks, from answering questions about campus resources to providing information on academic policies. Don't be afraid to use TritonGPT in different ways to get the most out of this helpful resource.

To get the most out of TritonGPT (or any AI model), you need to craft clear and well-structured prompts. Prompts serve as the cornerstone of communication with AI models. The best prompts precisely outline the task at hand, offer context to enhance comprehension, and include thoughtful follow-up questions and feedback, leading to the most accurate results. Take one of the training modules (in the drawer below) to help you get the most out of TritonGPT. 

Components of good prompts:

• Task (mandatory): always start with an action verb (Generate, Write, Analyze, Summarize, etc.) and spell out the end goal (single or multi-task requests)
• Context (important): Ask yourself 3 questions: What is the user’s background? What does success look like? What environment are they in?
• Examples (important): Examples help the model figure out what the final result should look like.
• Tone (optional, but nice to have): Ask TritonGPT to suggest a tone for the feeling you’re trying to convey (use professional but direct language.)

The key to using generative AI is to be specific about the task you’re trying to accomplish and to give enough information to the model so that it can respond in a way that’s meaningful and useful.

Module 1 - AI Foundations

This course covers foundational Artificial Intelligence concepts, UC policies regarding the usage of AI tools, and opportunities for application in Higher Education.  This self-paced course can be completed in approximately 30 minutes. 

• UC San Diego employees: Take the course in UC Learning with this link - Access the course.
• Not a UC San Diego employee?  Check out this version of our course, which was modified for external audiences: Access the course.

Module 2 - TritonGPT

This interactive review of TritonGPT provides foundational thinking when applying Language-based AI tools to processes.  Additionally, the course introduces the available Assistants and provides prompts you can try for each.   

• UC San Diego employees: Take the course in UC Learning with this link - Access the course.
• Not a UC San Diego employee? - This content is only available to UCSD staff. Email osi@ucsd.edu if you wish to learn more about the content.

Q. Where is TritonGPT hosted?  
A. TritonGPT is hosted on-premises in the San Diego Supercomputer Center, ensuring that our data is secure and controlled. 

Q. What version of Llama is powering TritonGPT?  
A. Within the TritonGPT platform, various Large Language Models can be leveraged to power different AI Assistants, each tailored to specific needs. For instance, the UC San Diego Assistant and General AI Assistant are driven by Llama 3, while JD Helper utilizes Llama 2. The reason for this distinction lies in JD Helper's unique workflow, which has yielded more favorable results with Llama 2 compared to Llama 3 in our tests.

Q. What is Danswer?  
A. Danswer is a company founded by two UC San Diego Alumni, Yuhong Sun and Chris Weaver. The Danswer technology integrates the large language model (LLM) with the UC San Diego context to produce natural language responses to anything users want to know about the campus.
 

Q. Is the model trained or affected by my queries or other information I enter?  
A. The information you place in your prompts will not affect the model. General users will not be able to access the information you enter. If you provide “thumbs” feedback, ITS can use that information to improve future responses.

Q. I have an idea for a new AI assistant. How can I get my assistant developed?  
A. The AI in Administration Workgroup can review and prioritize your request in collaboration with our development team. Review the workgroup information and contact the person who is most aligned with your work area.

Q. What if I have additional questions?  
A. We hope that your first resource for information will be TritonGPT.  Otherwise, you can email us at tritongpt@ucsd.edu Please be sure to check out our additional training, templates, and resources as well!

TritonGPT Example Videos - This series, Everyday I AI, is a collection of 3-minute videos that show example prompts and uses of TritonGPT.  New videos will be released gradually. Note: The first few videos introduce TritonGPT and Prompting.  Videos starting with #4 show practice prompts and responses.

TritonGPT User Interface Guide - An interactive look at what the various buttons do within TritonGPT.