google/embeddinggemma-2: one 768-dim space, uneven search

EmbeddingGemma 2 is Google DeepMind's open multimodal embedding model, positioned by the article as taking architectural advances from Gemma 4 rather than being a Gemma 2 variant. It sits in a different role from a large language model: instead of producing answers, it turns data into a 768-dimension vector that can be used for retrieval, similarity scoring, clustering, classification, RAG or recommendation. The article tested it on free Google Colab, where the assigned GPU was a T4, and the 740M-scale model loaded without trouble.

The experiment deliberately used only images with clearly documented licenses, self-written short Japanese texts and audio of known provenance, including LibriSpeech under CC BY 4.0 and Python-generated tones. The results showed that multilingual, cross-media retrieval is not equally strong in all directions. Japanese text queries such as asking for a cat photo or mentioning wanting coffee ranked the right text and image high, but using the cat image itself as a query did not cleanly surface the matching cat description, and a spoken-voice query did not pull up the actual speech clip within the top seven. A separate trial of Matryoshka Representation Learning, which can shorten the embedding to 128, 256, 512 or 768 dimensions, kept the same top-ranked space document at 768 and 256 dimensions in that one example.

The author also raises explainability as an open question: because different encoders all write into one shared representation, it is not obvious why an image and a text land near each other, and the cat-image result is a concrete case. Whether this model becomes practical for a personal cross-media RAG over lecture materials, photos, videos and recordings will likely depend on whether these direction-specific gaps narrow with larger, more varied data. The article notes that its results are as of October 7, 2026, and that Colab's GPU and the relevant libraries may change.