Google just leveled up EmbeddingGemma—now EmbeddingGemma 2 supports multi-modal embeddings, not just text. With a 740M parameter model, it can convert text, code, images, video, and audio into comparable vectors so you can literally do “content search with content.” The big win: it’s built for on-device use. Developers can run it directly on phones, laptops, and other terminals—meaning semantic search can happen locally without uploading your files to the cloud. That’s huge for privacy and latency. Under the hood, it maps different modalities into the same embedding space: the closer the meaning, the closer the vector. So you can describe a scene like “dog chasing a ball in a field” and retrieve relevant frames—even if there’s no matching text. It supports up to an 8K token budget, with sizable limits for audio/images/video, and adjustable embedding dimensions (768 down to 512/256/128) to save storage. Google also reports stronger code retrieval on MTEB Code tasks. What would you search locally first? #Embeddings #MultimodalAI #OnDeviceAI #SemanticSearch #VectorDB #CodeSearch
Want to learn more? Visit Explore the world, stay updated on travel insights and international affairs, and discover authentic stories from real life
评论
发表评论