Sign in Subscribe
4 min read

UIUC.chat Product Update #2

You’re receiving this email since you’ve signed up for UIUC.chat. We will only email you about product updates. This is our 2nd email in our two-year existence, but we have 1,700 git commits and 600 PRs across our frontend and backend. I hope this is ‘high alpha’ insight into the best uses of LLMs that we’re pushing into production. We are a small team of researchers and everything we build is free; enjoy!

Today we're launching:

  1. Tutor mode
  2. Deep Knowledge Control, an advanced prompting technique
  3. Analysis Page to visualize usage
  4. Anthropic Claude support
  5. Free Vision Language Model (VLM) hosted at our home, Illinois NCSA

Guided Learning, AKA "Tutor mode"

Infinitely patient without giving away the answer

💡
Bottom line: tutor mode prevents the AI from giving students the answer, and instead guides them to solve it themselves and offers hints along the way.

Many professors at Illinois use us as an AI teaching assistant. Now with a single click you can choose between the default behavior of 'be as helpful as possible' and the new Guided Learning mode, which won't give away answers. Instead, the AI will guide students through the steps by prompting a student's own problem solving. You can adjust these settings on your Prompting page.

Pro tip: you can leave your project on the default mode and only share a Tutor Mode Link - that link will always use tutor mode no matter what.

Deep Knowledge Control

Introducing a new way to control what your bot will know.

Limit your bot's world knowledge to only the documents you've uploaded, perfect for high precision tasks.

By default, the AI will prioritize knowledge in the documents you've uploaded. But in the case that none of the documents are relevant to the user's chats, the bot will fall back to its world knowledge from its training data. We do this to make all answers as helpful as possible.

Now, you can limit the bot to ONLY use the information from your uploaded documents. This can be helpful for things like analyzing contracts and searching inside long documents.

Unlimited Prompt Controls

Finally, to enable complete control, you can now bypass our internal prompt engineering. You can copy our standard internal prompt and customize it or write your own. That way, no matter what you're doing with your bot you can accomplish it on our website.

All prompt controls are on your project's Prompting page.

Analysis Page

The all-new Analysis page shows stats about your project's usage. If we're missing any stats you want to see, just reply to this email and we'll add those to the dashboard.

The analysis page, showing all the usage of one of our projects.

Concept Map of User Queries

The Concept Map visualizes all user queries made in this project.

A semantic similarity visualization of the chatbot's usage.

At the bottom of the Analysis page, we show a very cool visualization of your user queries. Similar questions are grouped closer together. Then we auto-generate labels for the groups we discover in your user's interactions. You can scroll around and review the transcripts for all your project's usage.

  • Similar topics cluster together
  • Different topics are positioned further apart
  • Common themes and knowledge gaps become visible

Learn more about semantic similarity visualizations.

Fun fact: we're launching these same maps to visualize your uploaded documents too, we just have to squash a few bugs first - these maps are tricky to constantly rebuild as new documents are added.

The smartest LLM from Anthropic is here

Claude Sonnet 3.5 offers the best LLM today, by most accounts, and now you can use the Anthropic API, with full support for images!

During your chat, upload any number of images and experience our advanced image understanding capabilities with Claude and GPT-4o.

Using images in chat.

Free LLMs hosted at NCSA

🧠
We’re thrilled to announce that Llama 3 support is now available for FREE to all users.

To access it, simply select in any chat - click the model name in the bottom bar. Naturally, there are sane usage restrictions to keep access fair for all users.

Select Llama 3 for free.

We self-host these models here at Illinois NCSA, and in the spirit of being a public university, I'm proud to share them publicly to give back to our community and to remove any barrier from people enjoying our app.

Coming soon: Qwen2 VL 72B - the largest and smartest Vision Language Model on the open source landscape today.

Please send me your feature requests 😅

We can’t wait to see what you’ll build with these new features! As always, your feedback is incredibly valuable. Just reply to this email with any questions or suggestions, don’t hesitate to reach out. I'm eager to build anything you need.

Member discussion