multi-qa-mpnet-base-dot-v1

Encodes text passages and queries into 768-dimensional dense vectors using MPNet architecture, enabling fast retrieval via dot-product similarity scoring. Trained on MS MARCO, StackExchange, and QA datasets to optimize for ranking relevance in information retrieval scenarios. Uses contrastive learning with in-batch negatives to align query and passage embeddings in the same vector space, allowing efficient approximate nearest neighbor search via FAISS or similar indexing.

Encodes queries and passages from multiple languages into a shared 768-dimensional embedding space trained on diverse QA datasets (Yahoo Answers, Natural Questions, TriviaQA, ELI5). The model learns language-agnostic semantic representations through contrastive learning across parallel and non-parallel QA pairs, enabling cross-language retrieval where a query in one language can retrieve passages in another. Architecture uses MPNet encoder with shared vocabulary across languages.

Encodes variable-length text sequences into fixed 768-dimensional vectors using mean pooling over token embeddings from MPNet's final layer. Supports efficient batching with dynamic padding to minimize computation on padding tokens, and includes optional attention-weighted pooling to emphasize semantically important tokens. Inference optimized for both CPU and GPU with ONNX export support for production deployment.

Produces embeddings compatible with FAISS, Pinecone, Weaviate, and other vector databases via standard float32 768-dimensional vectors. Embeddings are optimized for dot-product similarity (not cosine), enabling efficient approximate nearest neighbor (ANN) search using HNSW, IVF, or other indexing structures. Model outputs unnormalized vectors by default, which is critical for dot-product indexing performance.

Ranks candidate passages by relevance to a question using dot-product similarity between question and passage embeddings. Trained on MS MARCO, Natural Questions, TriviaQA, and ELI5 datasets where the model learned to align semantically relevant question-passage pairs in embedding space. Enables re-ranking of BM25 results or standalone ranking of pre-retrieved candidates without explicit relevance labels.

Extracts 768-dimensional contextual embeddings from text that can be used as features for downstream machine learning tasks (classification, clustering, similarity prediction). Embeddings capture semantic meaning learned from QA and retrieval training, enabling transfer learning without task-specific fine-tuning. Compatible with scikit-learn, XGBoost, and other ML frameworks via standard numpy/PyTorch tensor output.

Computes semantic similarity between arbitrary text pairs (sentences, paragraphs, documents) by encoding both texts and computing dot-product similarity between their embeddings. Similarity scores range from 0 to ~100+ (unnormalized dot-product) and indicate semantic relatedness regardless of lexical overlap. Useful for detecting paraphrases, duplicate content, or semantic equivalence without explicit training on similarity labels.

Exports model to ONNX and OpenVINO formats for deployment on edge devices, mobile platforms, and CPU-only infrastructure without PyTorch dependency. ONNX export includes optimizations for inference engines like ONNX Runtime, TensorRT, and CoreML. OpenVINO export enables deployment on Intel hardware with quantization support (int8) for reduced model size and latency.

Model weights available in safetensors format, a secure alternative to pickle-based PyTorch .pt files that prevents arbitrary code execution during model loading. Safetensors format is human-readable, supports lazy loading of individual weight tensors, and includes built-in integrity checks. Compatible with sentence-transformers, Hugging Face transformers, and other frameworks via safetensors library.