huggingface/sentence-transformers

State-of-the-Art Text Embeddings

Summary

Sentence Transformers is a Python framework for computing, training, and using text embeddings, rerankers, sparse encoders, and late-interaction retrievers. It exposes four model classes — SentenceTransformer, CrossEncoder, SparseEncoder, and MultiVectorEncoder — that share a common API and load from 15,000+ pretrained checkpoints on Hugging Face. The project is published under Apache-2.0 and the README states it "contains experimental software and is published for the sole purpose of giving additional background details on the respective publication."

What it is useful for

• Semantic search and semantic textual similarity across 100+ languages.
• Two-stage retrieve-and-rerank pipelines, where a fast bi-encoder retrieves candidates and a Cross-Encoder reranks them.
• Sparse retrieval with SPLADE-style models when you want inverted-index-friendly vectors instead of dense ones.
• ColBERT-style late interaction for higher-quality retrieval using token-level multi-vector representations.
• Fine-tuning on domain data. The README documents 20+ loss functions for embedding models, 10+ for rerankers, and 10+ for sparse encoders, with multilingual and multi-task learning support and evaluation during training.
• Multimodal embeddings and reranking for text, image, audio, and video through a single API.

How engineers can use it

The README recommends Python 3.10+, PyTorch 2.2+, and transformers v5.0+, installed with pip install -U sentence-transformers. For an editable dev install, clone the repo and run python -m pip install -e ".[dev]", pre-commit install, then pytest.

The four APIs follow the same pattern: load by Hub id, call encode (or predict for rerankers), then score.

• Dense embeddings: SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2").encode(sentences) returns NumPy arrays (README example shape (3, 384)); model.similarity(embeddings, embeddings) returns a cosine-style matrix.
• Reranking: CrossEncoder("cross-encoder/ms-marco-MiniLM-L6-v2") with model.predict([(query, passage), ...]) returns scores, and model.rank(query, passages, return_documents=True) sorts candidates. The README example correctly ranks Berlin-population passages by relevance.
• Sparse embeddings: SparseEncoder("naver/splade-cocondenser-ensembledistil") produces vocabulary-sized sparse vectors (README example shape (3, 30522) with ~99.84% sparsity). SparseEncoder.sparsity(embeddings) reports stats.
• Multi-vector (late interaction): MultiVectorEncoder("lightonai/GTE-ModernColBERT-v1") uses encode_query / encode_document to produce per-token 128-d vectors, scored via MaxSim.

For training, the README links to per-model overviews on sbert.net and to blog posts covering Matryoshka embeddings, static (CPU-friendly) embeddings, and binary/scalar quantization for post-training compression. The sbert.net docs cover installation variants (uv, conda, source), CUDA setup, and extras such as [image], [audio], [video], [train], [onnx], [openvino], and [dev].