TritonGPT Release Notes

TritonGPT introduces a redesigned sidebar with grouped chat history, assistant-specific threads, and a new “Explore Assistants” panel for enhanced discoverability. A streamlined assistant selection interface and prompt shortcuts improve usability and workflow. Users can now choose between multiple LLMs—GPT-4o and LLaMA 3.3 70B—based on task needs. Backend upgrades include a 128,000-token context window. The new User Settings panel enables theme toggling, auto-scroll, keyboard shortcuts, and full chat deletion. Legacy folders and dropdowns have been removed in favor of a cleaner, persistent navigation model. More details can be found here. 

Upgraded to the latest version of Onyx and improved the Related Questions feature. Users can now click on related questions to automatically populate them in the prompt window for a seamless experience.

Upgraded to the latest version of Onyx (formerly Danswer), integrating chat and search functionalities into a unified experience. The update also includes an upgrade to Llama 3.3 and the expansion of indexed websites—now totaling 63—to enrich the UC San Diego Assistant's contextual knowledge and improve its ability to answer a broader range of questions.

The JD Helper tool has been revamped to improve its functionality and reliability. Once you've developed your Position Overview, ask it to create a job posting, interview questions and other collatoral related to the hiring process. 

The Internet Search Assistant has been introduced, providing real-time informaton and answers.

A "Related Questions" feature has been added to the Fund Manager Coach, UC San Diego and Internet Assistants, providing thought-provoking topics and suggestions at the end of each response, tailored to the context of the original question.

Cleaner interface that allows for easier toggling between AI Assistants. 

Upload or paste emails, including key details like sender, subject, and content. This assistant will analyze them, provide a risk rating, and explain the ranking.

TritonGPT updated with a refreshed UI design and the ability to save and organize chats. Users can now upload and interact with documents using an AI Assistant. Additionally, users can search across all documents and websites, switch between Assistants within a chat flow, and access a Prompt Gallery to reuse previously created prompts. Other new features include the ability to see the context behind an Assistant's response, share chats with others, and interact with specific sources within the chat flow. A new AI Assistant, Fund Manager Coach, is now broadly available. View more detailed release notes.

TritonGPT is now available to all employees that are part of campus and health sciences, totalling close to 37,000 faculty, staff and student staff. If you are part of those units and are having difficulty accessing TritonGPT, contact the Service Desk. 

TritonGPT now uses the Llama 3 LLM to power the UC San Diego Assistant and the General AI Assistant. Llama 3 was trained on a much larger dataset (15T tokens) compared to Llama 2 (2.2T tokens), resulting in improved performance across various benchmarks. Llama 3 offers better reasoning abilities, code generation, and instruction-following capabilities compared to Llama 2.

As part of this enhancement, an upgrade was made to the Virtual Large Language Model (VLLM) environment. VLLM is a powerful and versatile open-source library that makes it easier and more efficient to use large language models (LLMs) for inference and serving. 

TritonGPT now uses campus Single Sign-on for authentication. This replaces the previous authentication method which was the Google Authenticator. This deployment also included some minor user interface improvements.

Enhancements and bug fixes to JD Helper: updated wording in the JD Helper flow to clarify what is needed in order to produce better outcomes in retrieving job cards. Enhanced JD Helper system prompt to provide more robust output. 

Launched production candidate that will support use by employees of campus and health sciences. Danswer enterprise code base and larger embedding model. User experience improvements. 

Added additional hardware into the environment in perparation for campus launch. 

Initial proof of concept version deployed and made available to over 400 pilot users.