Google just released Gemini Embedding 2, and yeah—this one is a big deal. It’s the first unified multimodal embedding model from Google that can take text, images, video, audio, and documents and embed them into the same vector space. Historically, you’d do annoying intermediate transformations (like speech-to-text for audio, OCR or image captioning for PDFs/images), and you lose semantic intent along the way—tone, urgency, background context, etc. With a unified embedding model, you preserve more of the original modality’s meaning, and that unlocks much better cross-modal retrieval. In the first part of the video I walk through a practical notebook: I embed text + images + audio together and show text-to-image search, image-to-image similarity, and audio-to-image retrieval. Then I zoom out into what I actually care about: building full-stack, agentic retrieval systems that go beyond “top-k chunks.” I show an agentic file search app that uses modality-specific chunking, stores embeddings in DuckDB, does clustering + classification + auto-tagging, and resolves cross-references at retrieval time so the LLM gets full context. Finally, because these systems make a lot of API calls, I show how I wired in Chargebee to handle plans, entitlements, and usage-based billing (e.g., 50 calls on free, more on pro). The model is still preview and pricier than older options, but if you’re building retrieval, clustering, or classification—this is worth testing.