Multimodal Embedding & Reranker Models with Sentence Transformers

The Sentence Transformers library now supports multimodal embedding and reranker models, enabling the encoding and comparison of text, images, audio, and video. These models map inputs from different modalities into a shared embedding space, facilitating cross-modal similarity search. This update allows for applications like visual document retrieval and multimodal Retrieval-Augmented Generation (RAG) pipelines. The process involves loading a model like Qwen3-VL and using the familiar API to encode various data types to compute similarities or rank documents.