PyLate

PyLate is an open-source Python library from LightOn, built on top of Sentence Transformers, for training and using ColBERT-style late interaction models. It covers fine-tuning on single or multiple GPUs, encoding queries and documents, indexing, retrieval, and reranking. The latest release listed is v1.6.0, and the library is MIT licensed.

What It Is

PyLate is a library for building and running late interaction retrieval models. It lets users construct ColBERT models from most pre-trained language models; when the base model is not already a ColBERT model, a linear layer is added to the encoder. Models are loaded and saved through the models.ColBERT class, and training uses the Sentence Transformers trainer.

Training Workflows

The docs show two training approaches. Contrastive training uses triplet datasets (query, positive, negative, with support for multiple negatives per query) and offers a configurable temperature, a cached contrastive loss based on GradCache to emulate larger batch sizes, and gathering across devices in multi-GPU setups. Knowledge distillation trains from the scores of a stronger teacher model using query, document, and score datasets. Hugging Face Datasets are supported, and a NanoBEIR evaluator is available for quick validation of English retrieval models.

Retrieval and Reranking

For retrieval, users encode documents and queries, add document embeddings to an index, and retrieve top-k results with scores. The default index uses FastPLAID. A WARP backend is available for XTR-trained models, and a TACHIOM backend provides CPU multi-vector retrieval. Without building an index, rank.rerank can rerank first-stage results using query and document embeddings.