Sentence Transformers

Sentence Transformers (also known as SBERT) is an open-source Python library for generating embeddings with models that run on your own machine. It grew out of BERT, the transformer that taught NLP to read text bidirectionally, but adds the training tricks needed to make whole sentences — not just tokens — comparable as single vectors. The library wraps thousands of pretrained models on Hugging Face, from tiny CPU-friendly workhorses like all-MiniLM-L6-v2 to large multilingual retrieval models.

This is your escape hatch from per-token API pricing. Local embedding means zero marginal cost, no rate limits, no network latency, and no data leaving your infrastructure — a hard requirement in many regulated environments. It’s also the path to fine-tuning: when generic similarity doesn’t match your domain (legal clauses, product SKUs, log lines), you can train a sentence-transformer on your own pairs, something closed API models won’t let you do.

Usage is three lines: pip install sentence-transformers, then model = SentenceTransformer("all-MiniLM-L6-v2") and embeddings = model.encode(texts). The encode call batches automatically and returns NumPy arrays ready for FAISS or any vector database, and built-in helpers like model.similarity() handle cosine scoring. You’ll typically prototype with a hosted API, then reach for sentence-transformers when cost, privacy, or domain fit says “run it yourself.”