Hey everyone!
These past two weeks, I worked on several major features and optimizations to get the plugin ready for user testing. First, I implemented the adaptive Auto-K selection algorithm (PR #38) that I mentioned in the last update. Instead of sweeping silhouette scores that often defaulted to K=2, the algorithm now calculates appropriate K values based on note count thresholds, giving consistent, high-quality clustering across both Native AI and local ONNX pipelines. Next, I added a resilient undo engine and smart auto-cleanup (PR #39), which tracks note movements in Joplin settings so users can revert all changes with one click, while automatically deleting empty sub-notebooks left behind. I also fixed UI text scaling and layout edge cases (PR #37), removed the K-Medoids pipeline (PR #40) to eliminate processing bottlenecks, and refactored the webview UI (PR #41) to streamline strategy previews and optimize the pipeline for 1,000ā5,000+ notes. With all these fixes in place, I published version v0.1.6 of the plugin!
Problem Faced: While testing the plugin on larger note collections (around 1k notes), I noticed that HDBSCAN creates a lot of noise points. Around 300 to 400 notes end up grouped under the "Uncategorized" cluster because density thresholds are currently too strict. K-Means, on the other hand, is working solid and ready to go.
Plan for the remaining two weeks is to fine-tune HDBSCAN parameters and noise reassignment so notes get grouped better, open a dedicated Joplin forum thread with a video demo to gather community feedback on v0.1.6, and address any bugs reported by users (currently working on recording the video).