In the context of Sentence Transformers, what is meant by a "bi-encoder" model?

A bi-encoder in Sentence Transformers refers to a model architecture that independently encodes two input sentences into fixed-dimensional vector representations, then compares these vectors to determine their similarity. This approach uses two separate neural networks (or two instances of the same network) to process each sentence in isolation. The resulting embeddings capture semantic meaning, and tasks like semantic search or clustering rely on measuring the distance (e.g., cosine similarity) between these vectors. For example, in a question-answering system, a bi-encoder might encode a user’s query and a database of answers separately, then rank answers by their similarity to the query’s embedding.

The primary advantage of bi-encoders is efficiency. Because each sentence is encoded independently, embeddings for large datasets (like product descriptions or support articles) can be precomputed and stored. This makes retrieval tasks fast at inference time, as comparisons reduce to simple vector operations. For instance, a search engine using a bi-encoder could precompute embeddings for millions of documents, enabling real-time similarity searches. Popular models like all-MiniLM-L6-v2 are optimized for this purpose, balancing speed and accuracy. Additionally, bi-encoders are easier to scale because encoding and comparison steps are decoupled, allowing distributed systems to handle each phase separately.

However, bi-encoders trade some accuracy for speed. Since the two sentences are processed separately, the model cannot capture fine-grained interactions between words in paired sentences. For tasks requiring deep contextual analysis (e.g., paraphrase detection), a cross-encoder (which processes both sentences together) often performs better but is slower. Developers choose bi-encoders when latency and scalability are critical. For example, a recommendation system might use a bi-encoder to match user profiles to products, while reserving cross-encoders for reranking top candidates. The choice depends on balancing performance needs with computational constraints.