Usage limits!

Went over the 10hr limit 😭

• 🙈 In this update, messages now count towards a single usage limit. The default is 3000 credits, which is equivalent to 3 million tokens of Gemini 3 Flash.
• 💸 Limits are enforced at the API level, and the chat level. Free models can still be used even when the limit is exhausted.
• 🔁 Limits reset every billing cycle, which is about 1 month from the same day. Built in mind to be extended with stripe in the future to be monetised.
• ❤️ This is a soft limit, so currently generating chats won’t stop halfway.
• 📋 Billing breakdown can be seen in the usage tab on account settings, will be extended in the future!

Other than that,

• 🍝 The notebook chat API and main chat API were merged, fixing some of the spaghetti code building up.
• 📁 The main chat can now reference ALL files too! That’s right, ALL FILES.
• 🤓 Notebook chats are still limited to files only linked to the notebook.

📹 Check out the awesome demo & screenshot i forgot to include in the video. Chat generation kinda slow, just skip forward.

Web search!

• Short devlog, but I added Perplexity Search as a tool to both standalone chat & notebook chat. 🔍
• 🐛 Took considerably longer cause I found some weird bug where thinking modes would become null on reload
• Check out the screenshot for the UI eyecandy! 🍭

Notebook & Chat updates!

Been really busy and I forgot to log :(

Notebooks

• ☁️ Notebooks now save! Changes made to the editor persist between reloads.
• ➗ Math equations can be edited using Latex. You can now create custom physics formulas :)
• ✨ Brand new animation when generating that unintentionally looks like siri (it’s kinda laggy though)

Chat

• 🤖Claude now works
• ⚒️ New UI to visualise the model’s tool calls
• 👀 Seeing which queries and sources the model pulls information from is now easier than ever
• 🌐 Plans to implement this for web search

Check out this amazing demo below 📹
(might be a lil blurry cause of the file limit)

More UI overhauls!!

I completely forgot to make devlogs 😭 Been really busy so far

In this update, I introduced a notion style notebook editor, allowing users to edit their AI-generated notes.
Along with that,

• I’ve polished up the library panel
• Chats and quizzes now open in a seperate tab
• Tabs can be dragged between the sidebar and the main panel, allowing for ultimate flexibility(check out the video!!!!)
• The main panel (notebook) can also be resized, making it more flexible on smaller devices like ipads.

In my next devlog, I’ll work on making edited notes save, and polish up the quiz UI (it was rushed from flavourtown)

Facelift!

• Made some huge design changes to notebooks! The interface is now tabbed, and notebook content is finally editable like Notion.
• Still ironing out formatting for the new editor, so math is a little broken
• New library interface is resizable and supports dragging!
  (Took longer than usual cause alot of things had to be moved around ❤️. Enjoy the videos!!)

RAG pipeline!

Finally migrated the old notebooks to a rag-based one. Most of the changes are backend, so theres not much to show :(

• Files (PDFs, slides, lecture notes) were directly used as sources for notebooks before. This was slow, inefficient, and wasted tokens. Now, files are converted to markdown on upload.
• This markdown is then processed, and broken down into chunks in the rag pipeline, making chat faster and more accurate!
• Additionally, chat uses the new AI gateway/Openrouter provider to ensure fallback and consistent speeds regardless of provider downtime (plagued the app during flavourtown)

On preliminary testing, generation speeds jumped from 30s ttft to 10s ttft, with chat seeing the most improvements.

The next update should improve UI and refine subject prompts (a lil too robotic rn)!

Added a model selector to the chat menu! Along with this, thinking modes were standardised across all models, and usages are now tabulated.

I plan to add tools too, so the chat window can actually retrieve information from notebooks or materials uploaded.

Stardance!! 🌟

.

Back story

This project is a continuation from flavourtown (https://flavortown.hackclub.com/projects/10428), initially to help students use AI to make studying faster.

A short summary for those unfamiliar: Mesa AI would create summary notes, quizzes and chats in a interface called notebooks. It’s meant to be similar to Notebook LM, but actually make notes and offer more exam accurate (subject-based instructions) notes & quizzes.

Unfortunately, I couldn’t finish Mesa AI in time (took almost 100 hours!). In Stardance, I aim to complete this project’s ambitious plans, and finally ship it.

Here are the things I intend to change or add:

• Complete all the unfinished pages
• Migrate to a fully custom RAG engine
• Move towards openrouter and Vercel AI gateway after that gemini outage which plagued Mesa AI in most of flavourtown
• Move from a read-only summary notes to notes which you can actually edit, and Mesa AI will edit them with you (agentic studying)

.

All those changes would need me to fundementally rewrite how Mesa AI was built, and make this journey on Stardance even more interesting. Mesa AI could also evolve to not just be a notetaker, but a researcher and an entire online tutor too.

.

First Devlog

To start off this rewrite, I’ve already made the custom RAG engine, which replaces Google’s file search API (too slow, unreliable, and locks me into using a Google API key).

All files are indexed automatically by a combination of Jina (PDF to markdown) and Mistral (PDF to markdown too), along with OpenAI’s embedding models. A 20-word summary is then generated with Gemma 4 such that search queries are more relevant without relying on file name alone.

To test this, I made a barebones boilerplate to simulate a real chat. (See the image!) Sorry I don’t have more pictures! This devlog was mainly backend.

In my testing, indexing (turning the file into embeddings and the summary) takes about 20s, with subsequent messages regarding that file taking about 5-8s (search and retrieval). Really pleased with the accuracy though!

Sadly i forgot to install hackatime on my mac so a few hours were lost when i was working outside :(