quaterion

Why we included this project

Getting embeddings that actually separate your data is usually the hard part of a similarity system, not the vector store itself. Quaterion is a training framework built for that last mile: it fine-tunes pre-trained models into similarity-aware encoders with contrastive and metric-learning losses, so you can specialize a general model to your own domain without training from scratch. Because it sits on PyTorch Lightning, you get distributed training and checkpointing out of the box, and a built-in caching mechanism keeps training fast even on a laptop GPU. The framework is deliberately open to redefinition, so you can swap in your own encoders and head layers, and the docs include tutorials for NLP and computer-vision use cases. If you already run Qdrant or another vector database, it is a natural companion for producing the embeddings those stores serve.