dérive

dérive is a recommender over concept embeddings that suggests rare concepts close to or far from your taste, learned by Rocchio relevance feedback. This repository runs the recommender of the live demo in a terminal.

Quick start

git clone https://github.com/gustavecortal/derive.git
cd derive
node read.ts

Type an option's number to read it, s to save the card, d to dislike it, and q to quit.

How it works

Each of the 2,028 concepts is embedded from its term and description with Qwen3-Embedding-4B. Under each card, Closest is the unread concept most similar to it. Other field is the one most similar to your taste, outside the fields of the card and of Closest. Far is drawn among the concepts least like the last ten you read. Your taste is the mean embedding of the concepts you picked or saved, minus half that of those you disliked.

Rebuild the embeddings

uv run embed.py

This rewrites data/vectors.bin from data/concepts.json. The model is an 8 GB download and runs in under a minute on a GPU, or over half an hour on a CPU.

License

The code is under the MIT license. The concepts are under CC BY-SA 4.0, since many descriptions are adapted from Wikipedia.