One model, one 768-dimensional space, three LanceDB indexes: search 10,000 VGGSound clips with a sentence, a soundtrack, or a recording of your own, in a few lines of Python or a few clicks.

You have 10,000 video clips and a sound stuck in your head. A church bell, maybe. The clips have no labels for it, or the labels are wrong, or the thing you want lives in the audio track that no image model ever looked at.

Searching media by meaning has usually meant stitching models together: one for frames, one for audio, one for text, and a vector store for each. Then you compare scores across models that were never trained to agree.

EmbeddingGemma 2 , released by Google DeepMind on October 6, 2026, removes the stitching. One 740M-parameter model maps text, images, video, and audio into a single 768-dimensional space.…