Doze Sort: let a local model file the notes, and keep the originals

This submission for the Hacktoberfest Open-Source AI Challenge focuses on a tool called Doze Sort. The application helps users organize notes with filenames such as note-03.md. To use Doze Sort, an open-weight model compares the contents of each note with topic descriptions and generates a plan. The user can then review and execute that plan, resulting in organized copies of the notes in the suggested folders.

If the score is low or the result is uncertain, the note goes to a _review folder. Identical files are kept as a single canonical copy. The original notes remain untouched, ensuring the user retains their work.

Doze Sort is designed for small English Markdown or text inboxes, helping users file their notes more efficiently. It can be accessed via a browser demo or a command-line interface (CLI). The browser demo uses fictional files included in the repository, while the CLI accepts user-chosen inboxes and topics. Doze Sort never scans other folders in the background and runs on a loopback address.

The app was developed using MiniLM-L6-v2 as the content classifier, with Apache-2.0 open weights and pinned revision. Sentence Transformers runs the model on CPU, producing 384-dimensional embeddings for each note and topic description. Cosine similarity is used to assign a topic, requiring a top score of at least 0.30 and a top-two gap of at least 0.06. These thresholds are set as heuristics and not probabilities.

The copy engine checks the source filenames and hashes before applying the plan, ensuring that the user's original notes remain unaltered. If the inbox changes, a fresh plan must be created. Undo functionality is available to move completed runs into recovery, retaining all files and preventing any loss. The app includes seven behavior and integration tests, covering various scenarios such as stale input, duplicate references, replay prevention, and recoverable undo.

The project is licensed under MIT, and the original notes, tests, and recorded results are included. The examples cover cooking, gardening, and travel, all fictional original material. No client or employer files, credentials, dependencies, or model weights are committed to the repository. The app demonstrates the benefits of open innovation by classifying notes without an internet connection or per-file API charges.

The text stays on the computer running the model, and the app has no analytics, remote fonts, or background folder watchers. The policy is inspectable, allowing users to change topic descriptions or score thresholds and rerun the example inbox.