IT Center Changes

Features:

• Support for new models:
  • gpt-oss-120b and gpt-oss-20b
• Chat export function added

Bugfixes and Improvements:

• Improved display of formulas
• Document content is now sent to LLMs as a user message (previously system)
• Strict role sequence in prompts (no two consecutive messages with the same role)
• Improvements to timeouts and retries
• Various design optimizations
• The correct time period is now displayed when message limits are reached
• Title generation with Mistral Small now works correctly
• The sidebar can now be expanded and collapsed again

Features:

• Newly supported models:
  • gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-chat
  • Beta support for various models from the Llama, Mistral, Gemma, Qwen, Teuken, and InternVL families
• Users with access to multiple user groups can now switch between groups in the frontend
• Added notifications for automatically deleted chats, along with a new modal for manually deleting or extending these chats

Bugfixes and Improvements:

• Content pages are now displayed correctly across all tenants
• Optimized print layout for chat history

Language models (new for employees):

• OpenAI GPT 4.1
• OpenAI o4-mini
• OpenAI o3

Language models (new for students):

• OpenAI GPT 4.1
• OpenAI o4-mini

Bugfixes and Improvements:

• Summer theme revised
• Improvement in cost tracking: Prompt caching and various error cases are now taken into account
• The chat deletion dialog now always shows the correct chat title
• File upload button is reactivated after a file that has already been uploaded is removed

Features: 

• Share chats: Chats can now be shared with others

Language model:

• Llama 3

Bugfixes and Improvements:

• When loading branched chats, the last option is now always displayed first
• Used tokens are, if possible, taken directly from the chat responses and no longer calculated separately
• Added info text for slow chat responses
• General design optimizations
• Improved display of formulas
• Improvements in the admin panel:
  • API key of access data can no longer be viewed after adding
  • Reporting functions have been improved
• Search function: Errors when quickly changing the search word have been fixed
• Reasoning models:
  • Title generation now works without errors.
  • Code and formulas are displayed correctly in reasoning models
• Images can now be selected correctly when uploading files
• Bugfixes in the admin panel:
  • Help page entries are no longer deleted when saving general settings

Features:

• Support for models hosted directly by OpenAI
• Multi-client capability
• Added chat search in the sidebar

Language models (new for students):

• OpenAI o1
• OpenAI o3-mini

Bugfixes and Improvements:

• Optimized display of the prompt input field
• Improved display of the editing of prompts
• Preparations for OpenSource models
• User initials now always use the correct font
• An & within generated code is no longer output as &;

Bugfixes and Improvements:

• Fixed a bug in the title generation for the chat history for o1 models

Bugfixes and Improvements:

• Accessibility improvements
  • Page language set correctly
  • Integration of semantically labeled headings, attributes and other HTML elements
  • Contrast values improved
  • Improved use with keyboard
• If errors occur, users are no longer redirected to a separate error page
• Line breaks in prompts are now displayed correctly in the chat history
• Fixed an error when uploading files for o1 models

Features:

• Access to RWTHgpt is now also available for all RWTH students. More information can be found in the corresponding blog post.
• Access for students is linked to active status student RWTH status.

Language models (for students):

• GPT-4o
  • 6,000 tokens per week
• GPT-4o-mini
  • 25,000 tokens per week

Features:

• From now on, access to RWTHgpt no longer requires a role in RWTH Identitiy Management.
• Access is now linked to the active status of employee RWTH and FB 10 (in the context of UKA).

Language models:

• GPT-o1-preview is now available
  • 25 interactions per hour
• GPT-o1-mini is now available
  • 100 interactions per hour with slightly limited functionality
• These models first create a long “chain of thought” internally before responding. The answers are more complex and focused, but the generation can also take longer.
  Recommended for use in programming, mathematical or scientific contexts that require complex reasoning.

Bugfixes and Improvements:

• Improved response behavior
• Performance optimization
• Improved readability in the course of accessibility