Google just dropped Gemini Embeddings 2, a new model that natively understands images, videos, and text all at once. In this video I use it with Claude Code to build a full visual search engine from scratch.
The crazy part is you don't have to build any of the chunking or ingestion pipeline yourself anymore. Just describe what you want, point it at your files, and Claude Code handles everything. It extracts content from images, generates descriptions, builds out your Pinecone vector database, all of it.
You basically just throw everything you want to be searchable at it and it works. This is a massive unlock for anyone building with RAG.