Sentence-Transformers
Sentence-Transformers is a library for using and training embedding and related text-matching models. It supports representations for sentences and passages as well as cross-encoder scoring, making it useful for retrieval experiments where model choice, input formatting and task-specific evaluation matter more than a generic notion of semantic similarity.
What it is
A sentence-transformer model encodes text into a vector that can be reused for similarity search, clustering or other tasks. The library also exposes cross-encoders, which jointly process a pair and produce a score rather than independently reusable embeddings. Models differ in training objectives, languages, maximum input lengths and required query or document instructions. The library is therefore an implementation toolkit, not one fixed embedding model. A retrieval system typically pre-encodes documents, encodes queries at runtime and optionally uses a cross-encoder to reorder the retrieved candidate set.
What the work involves
The practitioner selects a model suitable for the language and task, checks its model card and documents preprocessing and normalization. Retrieval evaluation uses representative query–passage judgments; fine-tuning uses defensible pairs and negatives with a separate test set. Batching and device selection are measured for ingestion and query workloads. Useful artifacts include the encoding code, model version, evaluation results and saved training configuration. Changing models requires rebuilding document embeddings and checking score thresholds, even when the output vector dimension happens to remain the same.
Illustrative example
A multilingual help center tests several supported models on queries in the languages its users actually write. The team checks whether an informal question retrieves the right article and whether exact product codes are preserved. It then uses a cross-encoder for the best initial candidates and compares the full pipeline with embedding retrieval alone. A model that performs well on one language but loses relevance in another is not selected solely because its overall public benchmark result is high.
Limits and common mistakes
The library cannot guarantee relevance for every supported model or task. Truncation can remove the decisive sentence, training data may not cover specialist vocabulary and similarity scores are not calibrated confidence. Cross-encoders also add per-candidate computation. Good use requires model-specific documentation, representative relevance labels and operational measurement. Framework convenience should not hide which encoder, objective and input format produced the vectors or scores used by the application.
Prerequisites
Related skills
- → is an instance of: Embedding Models
Sources and further reading
- Sentence Transformers usage
Official guidance on model use, embeddings and similarity computation.
- Retrieve and rerank
Explains bi-encoder retrieval and cross-encoder scoring as distinct library capabilities.
Last updated: 2026-10-10