Skip to main content

System Status: 

April 2, 2025 Release Details

Summary of Updates to TritonGPT as of April 2, 2025

Overview

This release introduces a redesigned navigation and assistant selection interface for TritonGPT, improving accessibility, task flow efficiency, and visual clarity. Additionally, it includes significant backend upgrades such as expanded context windows and support for multiple large language models (LLMs).

What's New and Changed

Left Sidebar Redesign

  • A new "Assistants" section appears at the top of the sidebar, replacing previously ungrouped assistant chat folders. Users can now switch between assistants (e.g., UC San Diego Assistant, General AI Assistant, Job Description Helper) using a more intuitive, selectable list.
  • Once enabled in the user settings (top right), a new "Prompt Shortcuts" button appears below "New Chat", enabling quick access to reusable, task-based prompts.
  • You can now also search through your previous chats by hovering on the "Chats" title in the sidebar and clicking the search icon
  • Chat history is categorized by recency (e.g., "Today," "Previous 7 Days") for improved retrieval.
  • Assistant-specific chat threads are nested within their respective assistant headings, reducing clutter.

Enhanced Assistant Discoverability

  • An "Explore Assistants" option is now available beneath the main list of assistants to help users discover additional tools such as Fund Manager Coach, Internet Search Assistant, and Email Phishing Analyzer.
  • The Explore Assistants panel provides a searchable and filterable interface where users can browse featured assistants. Filters include Pinned, Mine, Private, and Public.

Language Model Selection

  • Users now have the ability to select between different LLMs depending on their use case. This enables tailored performance based on model capabilities and context requirements.
  • Users can select from multiple large language models (LLMs) to tailor performance to specific tasks. The supported models and their optimal use cases are:
    • GPT-4o: Developed by OpenAI, GPT-4o is a multimodal model capable of processing text and images. It excels in text-based applications such as data analysis, real-time language translation, and writing.
    • LLaMA 3.3 70B: Released by Meta, LLaMA 3.3 70B is optimized for multilingual dialogue and instruction-following tasks. It performs well in coding assistance, multilingual communication, and is efficient and capable of using the "Agentic Search" feature.

Details about the hosting and use of these models is found on TritonGPT Large Language Models

Message Composition Area

  • The assistant model in use (e.g., GPT 4.0) is now displayed beneath the input field.
  • TritonGPT now supports a significantly expanded context window of 128,000 tokens (approximately 96,000 words). This allows users to process and analyze larger documents, extended chat histories, and complex datasets without loss of context.

User Preferences Panel

  • A new User Settings panel is accessible from the top-right menu, allowing users to customize their experience.
  • Users can toggle between light and dark themes.
  • Auto-scroll can be enabled or disabled to control whether the chat view automatically scrolls to new messages.
  • Keyboard-based prompt shortcuts can be toggled on or off.
  • A red "Delete All Chats" button allows users to permanently clear all chat sessions. This action is irreversible and helps support data privacy and workspace management.
Contact wts@ucsd.edu.