jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers
The episode discusses the "jina-embeddings-v5-omni" paper, which introduces the GELATO method for creating multimodal embeddings. By using frozen vision and audio encoders connected to a text model via small projectors, researchers preserve existing text knowledge and achieve high efficiency, allowing for interleaved text, image, and audio processing.