Product releases
Explore Onyx-powered features and other TritonGPT capabilities after UC San Diego deploys and verifies them.
Feature history
Search deployed product changes by year and feature area.
Platform and models
TritonGPT upgrades to Onyx v3.3. Deployed features include structured indexing for spreadsheets and CSV files, smoother chat streaming with live LaTeX rendering, queued messages, image extraction from PDFs, and connector improvements. System hardening and observability work also improved performance and reliability.
TritonGPT
Adds GPT 5.4, Gemini 3.1 Pro and Flash, Gemma 4 26b, and Claude Sonnet 4.6, and retires the OpenAI 4 Series. The UC San Diego Assistant gains Parking and Busyness tools, the Internet Search Assistant moves to a more reliable web search, and answer quality and safety have been improved across the platform. Read the April 14 release details.
TritonGPT
This release refreshes the chat interface and organizes Documents and Chat Folders into Projects. It also adds GPT-5.1, Gemini 3 Pro, Deep Research, and personalization options. Read the November 24 release details.
Platform and models
OpenAI's GPT-5 model is available for all AI Assistants.
Platform and models
GPT OSS, OpenAI's open-weight model, now powers TritonGPT by default. Models are served through a centralized API service for consistent management. The new Data Dictionary Assistant helps report developers navigate data available in the Activity Hubs.
Platform and models
An improved embedding model increases response accuracy. All models are now served through LiteLLM for more efficient and scalable model management.
Users can also change their default model in User Settings.
TritonGPT
Paste a meeting transcript to generate a detailed summary with action items.
Platform and models
GPT-4.1 is available for all AI Assistants. It supports text and image inputs and brings updated coding, reasoning, and instruction-following capabilities. This release also includes fixes for My Documents.
Platform and models
Llama 4 Scout replaces Llama 3.3 as the default model.
TritonGPT
Users can create folders and upload documents to organize source material and generate new content from multiple files.
Campus assistants
The UC San Diego Assistant now evaluates prompts to determine when UC San Diego context applies. The change reduces irrelevant campus context when the General AI Assistant is a better fit.
TritonGPT
TritonGPT introduces grouped chat history, assistant-specific threads, and an Explore Assistants panel. Users can choose between GPT-4o and Llama 3.3 70B. Backend upgrades expand the context window to 128,000 tokens. User Settings adds theme selection, auto-scroll, keyboard shortcuts, and full chat deletion. Read the April 2 release details.
TritonGPT
TritonGPT upgrades to a newer Onyx release and improves Related Questions. Users can select a related question to add it to the prompt window.
Platform and models
TritonGPT upgrades to Onyx, formerly Danswer, which brings chat and search together. The release also adds Llama 3.3 and expands the number of indexed websites to 63.
TritonGPT
JD Helper adds support for creating job postings and interview questions from a Position Overview. Internet Search provides current web information. Related Questions suggests follow-up topics in the Fund Manager Coach, UC San Diego, and Internet Search assistants.
TritonGPT
A cleaner interface makes it easier to move between AI Assistants.
TritonGPT
Users can upload or paste an email for analysis. The assistant returns a risk rating and explains the factors behind it.
TritonGPT
TritonGPT adds saved chats, file uploads, document search, assistant switching, a Prompt Gallery, shared chats, and source interaction. Fund Manager Coach also becomes available. Read the August 13 release details.
Platform and models
TritonGPT adds Llama 3 for the UC San Diego Assistant and General AI Assistant. The vLLM environment also receives an upgrade to serve models more efficiently.
TritonGPT
JD Helper receives fixes and clearer instructions for retrieving job cards. Its system prompt also changes to produce more consistent output.
Try a broader search, another year, or all areas.