Why your RAG pipeline probably needs three encoders, not one

What it does

Sentence Transformers is a Python framework that wraps transformer networks into three distinct encoder APIs: dense bi-encoders (SentenceTransformer), cross-encoder rerankers (CrossEncoder), and sparse encoders (SparseEncoder). It gives you access to over 15,000 pretrained models from Hugging Face—including MTEB leaderboard entries—and supports fine-tuning all three architectures with task-specific losses for semantic search, similarity scoring, and paraphrase mining.

The interesting bit

The library does not treat embeddings as a commodity. It splits the retrieval problem into fast candidate generation, slow precise reranking, and lexical sparse matching, then exposes training pipelines for each with 20+ losses for embeddings and 10+ for rerankers and sparse encoders. It also quietly extends the same API to image, audio, and video inputs, though you will need to dig past the text quickstarts to find them.

Key highlights

• Three-model API: SentenceTransformer, CrossEncoder, and SparseEncoder cover dense, reranker, and sparse (e.g., SPLADE) representations.
• Massive model zoo: over 15,000 pretrained models on Hugging Face, including MTEB leaderboard entries.
• Training suite: 20+ losses for embeddings, 10+ for rerankers and sparse encoders, with evaluation during training and support for BERT, RoBERTa, XLM-R, BART, etc.
• Multimodal and efficiency extensions: supports image, audio, and video inputs; companion blogs detail Matryoshka embeddings, static embeddings, and quantization.
• Optional extras for ONNX, OpenVINO, and hardware-specific deployments.

Verdict

If you are building search, recommendation, or RAG pipelines and need more than a single embedding model—especially if you want to fine-tune dense, sparse, and reranker components from one library—this is the default toolkit. If you only need a single off-the-shelf embedding API without training or reranking, it may be more machinery than necessary.