paraphrase-MiniLM-L6-v2

Generates fixed-dimensional dense vector embeddings (384 dimensions) for arbitrary text sentences using a distilled BERT architecture (MiniLM-L6) fine-tuned on paraphrase datasets. The model encodes semantic meaning into continuous vector space, enabling similarity comparisons between sentences without explicit keyword matching. Uses mean pooling over token embeddings and applies layer normalization to produce normalized vectors suitable for cosine similarity operations.

Computes pairwise cosine similarity scores between sentence embeddings using normalized dot-product operations. The model's output vectors are L2-normalized, enabling efficient similarity computation via simple dot products (avoiding explicit cosine formula overhead). Produces similarity scores in the range [-1, 1], where 1 indicates semantic equivalence and negative values indicate semantic opposition.

Processes multiple sentences in parallel batches through the MiniLM encoder, applying mean pooling over token-level representations to produce sentence-level embeddings. The sentence-transformers library handles batching, padding, and attention mask generation automatically. Supports configurable batch sizes and pooling strategies (mean, max, CLS token), optimizing throughput for CPU and GPU inference.

Provides the same semantic embedding capability across multiple serialization formats (PyTorch .pt, ONNX, OpenVINO IR, safetensors) and inference engines, enabling deployment in diverse environments without retraining. The model can be exported to ONNX format for cross-platform inference, quantized for edge devices, or compiled to OpenVINO for Intel hardware optimization. Sentence-transformers handles format conversion and runtime selection automatically.

Enables semantic search by embedding both queries and documents, then ranking documents by cosine similarity to the query embedding. Unlike keyword-based search, this approach captures semantic intent (e.g., 'car' and 'automobile' are similar) without explicit synonym lists. The model is specifically fine-tuned on paraphrase pairs, making it particularly effective for matching semantically equivalent but lexically different text.

The model is compatible with text-embeddings-inference (TEI), a specialized inference server optimized for embedding models. TEI provides a REST API for embedding generation with features like batching, caching, and automatic GPU optimization. This enables deploying the model as a microservice without writing custom inference code, supporting horizontal scaling and load balancing.

While primarily trained on English paraphrase data, the model can process non-English text and compute cross-lingual similarities due to BERT's multilingual subword tokenization. However, performance degrades significantly for non-English languages because the paraphrase fine-tuning was English-only. The model tokenizes non-English text into subword units and produces embeddings, but semantic quality is substantially lower than for English.