AnglE

Why we included this project

AnglE is a compact library for training and running sentence embeddings, the vectors that power semantic search, retrieval-augmented generation, and text similarity. You can train embeddings from BERT-style models or from LLMs like LLaMA, Mistral, and Qwen with only a few lines of code, and it also works as a general inference framework for loading other transformer-based embedding models. Teams that need vectors tuned to their own domain data, rather than generic off-the-shelf models, will get the most from the training side, which supports several contrastive losses and runs on one or multiple GPUs. The method is described in a peer-reviewed ACL 2024 paper, so the approach is documented rather than a black box, and pretrained models are available for anyone who just wants to plug in and go. If you are building retrieval or similarity features and want control over how your embeddings are trained and served, this is a practical starting point.