Inspiration
Many of students note live on physical or digital notebook then ended as fosil, they won't touch it again (thats happen to me too at college). But when need the notes hard to find right? NotebookLM or obsidian it's too complicated for students.
What it does
Ritize! convert students note to the vector database use BGE-M3 model. Student can chat with own notes such as: "notes about Shipaton 20 days ago".
Using AI SDK + Gemini for reasoning, the student query send to MCP to access the notes from their own databases. Other students notes can't accessed.
The side effect user can also track their study times. Such as learn AI = 340 hours, learn Economic = 130 hours, etc.
How we built it
I build with Supabase for the database, AI SDK for client side, Django for the ORM. Notes chunking use Stanza and BGE-M3 as embedding model. Mobile application use IONIC + Capacitor + React.,
For the uploaded notes and canvas, I need extract content inside it and use batching method with Gemini Multimodal to decrease token usage. The results is text to be convert as vector with BGE-M3.
Challenges we ran into
The challenge is token optimization. Memory implementation for the LLM it self is challenging, to many context increase token usage - too small context break accuracy. For that I use "overlap strategy", this method prevent LLM lost memory too much but keep token small as possible.
Accomplishments that we're proud of
Success to make students note from 3 input method (text, canvas / stylus and upload notebook photo) retrieve with chat very accurate in any language. Such as in the note says: "cabe" but user query with "chili" the result is very accurate in the context level.
What we learned
AI: if we can optimize, the cost can very small but the result still amazing.
What's next for Ritize! Power Up Study Notes
For next month the next features is "convert notes to the voice" so the students can hear it in the road any time.
Log in or sign up for Devpost to join the conversation.