A local-first AI notes app where thoughts, files, images, and code collapse into a searchable semantic galaxy.
No folders. No tagging. Just search by vibes.
demo.mp4
VoidDrop is a desktop application built with Rust, Tauri, and React. At its core is a Black Hole — a drop zone where you feed it anything: a fleeting thought, a PDF, an image, a link, a code snippet. VoidDrop converts everything into AI embeddings and stores them locally, letting you retrieve information by meaning rather than exact keywords or folder structure.
You don't organise. You don't tag. You just search by vibes.
I'm the kind of person who dumps everything into Notes, reminder apps, or onto the desktop and hopes future-me can find it later.
So I made VoidDrop – a place where I can do exactly that, except future-me can now retrieve things in seconds instead of digging through chaos.
Head over to the Releases page and download the latest build for your operating system.
When you open the app for the first time, it may take a minute to download the models. After that, the models will load much quicker and not require an internet connection.
Important Note for macOS Users
Because VoidDrop is an independent project without an official Apple Developer Certificate, macOS's Gatekeeper will strictly flag the app as "damaged" and refuse to open it. The file is perfectly safe, it just needs its quarantine flag removed.
After dragging the app into your Applications folder, run this command in the terminal:
xattr -d com.apple.quarantine /Applications/VoidDrop.appNote: For macOS, VoidDrop is compiled exclusively for Apple Silicon (M1 chips and newer).
Legacy Intel Macs are not supported – the underlying ONNX Runtime no longer provides prebuiltx86_64-apple-darwinbinaries, and Rust itself has demoted the target to Tier 2.
- Backend: Rust + Tauri
- Frontend: React + TypeScript + TailwindCSS + Framer Motion + Three.js
- Storage: SQLite with
sqlite-vecfor vector storage and cosine similarity queries - Models: Local sentence embedding model (384d) + CLIP (512d) for images
The Latent Space and fragment list use performance-conscious rendering:
- Three.js + HTML5 Canvas for the galaxy (avoiding hundreds of DOM nodes)
- Virtualised lists so only visible cards are in the DOM
- Lazy loading for file and image previews — content renders only when a card is expanded
Drop anything into the void:
- Text Notes: Typed directly
- Documents: PDF, DOCX, PPTX, and more
- Images including HEIC, with automatic conversion
- Code: File type and language are detected and passed to the model for better context
Search by meaning, not keywords. Powered by local embedding models, results are ranked by cosine similarity.
- Text uses a 384 dimensional sentence embedding model
- Documents use the same model as for thoughts, and are chunked automatically to find the average meaning
- Images use CLIP, which shares an embedding space with text — enabling cross-modal search. Raw CLIP scores are calibrated so images rank accurately alongside text results
Every fragment the user has ever saved is plotted as a star in a galaxy using Orthogonal Projection.
Two perpendicular vectors define a 2D plane in the 512 dimensional embedding space. Each fragment's position is computed by projecting its embedding onto these axes using dot products.
- Cheap: Only requires dot products, no iterative fitting
- Absolute: Independent of the data; pure linear algebra
- Consistent: The projection vectors are generated once and never change, so points don't shift between sessions.
The result is a unique, personal galaxy whose shape is determined entirely by the semantic structure of your own data.
This project is licensed under the MIT License