Releasing my AI note app built as a side project after 3 months
I built an AI note app as a side project. Authentication is simple with Google, and I added a feature where AI summarizes note contents.
Here's a summary of the major things I handled during development.
1. Note search using vector embeddings
- I embedded all note contents to enable similarity-based search. Contextual search is definitely more natural than keyword search, so I'm satisfied.
2. AI summarization processed on-device
- At first I used a cloud API, but it was costly and slow, so I switched to a small model that runs locally. Response speed actually got faster, and it works offline now.
3. All data stored locally
- Given the nature of notes, I thought privacy was important, so all data is stored only on the device. I added an export-to-file feature so you can back up on a PC without cloud sync.
There are still many shortcomings, but I plan to improve it step by step. If anyone tries it and gives feedback, I'd really appreciate it.
I'll leave the link in the comments.
9 answers
Agree, I really needed an AI note summary, and this is well done.
Switching to on-device seems like a good choice. I initially used the cloud API too, but got shocked by the charges and moved to local. However, with embedding search, index management becomes crucial as the scale grows. By the way, how did you split your chunk size?
Oh, I didn't know this. That's interesting.
Well, I think it's a bit dangerous to assume that privacy is guaranteed just by storing data locally. If you lose your device or delete the app, won't all your data be gone? If the only backup is exporting files, that could be inconvenient.
Haha, I'm also thinking about making something as a side project, so I'll refer to this.
I made a similar note app about six months ago too, and on-device AI summarization had worse thermal issues than I expected. Would a smaller model help? Anyway, I like that it's stored locally. If you have a channel for feedback, please let me know.
Nice lol, at least it has file export, so that's better.
I've been having similar concerns, and it's impressive that you switched to on-device. Could I ask which local model you used? Also, I'm curious whether the summary quality varied depending on the length of the note content.
The content is good, but I find it hard to agree that vector embedding search is always better than keyword search. When looking for a specific date or exact phrase, keywords are much faster and more accurate.