Gemini Embedding 2 is now generally available via the Gemini API and Enterprise Agent Platform. It is the first Gemini embedding model to unify text, images, video, audio, and documents into a single semantic space across 100+ languages. Key capabilities include interleaved multimodal input processing in a single API call, task prefixes to optimize embeddings for specific use cases (question answering, code retrieval, clustering, classification), Matryoshka Representation Learning for dimensionality reduction (3072 down to 768 dimensions), and a Batch API offering 50% cost reduction. Real-world adopters include Harvey (3% Recall@20 improvement for legal search), Supermemory (40% Recall@1 improvement), and Nuuly (visual search accuracy from 60% to 87%). The post includes code examples for building agentic RAG pipelines, visual search, reranking, clustering, and anomaly detection, with integrations for Pinecone, Weaviate, Qdrant, and ChromaDB.
Table of contents
About Gemini Embedding 2Agentic retrieval-augmented generation (RAG)Multimodal searchSearch rerankingClustering, classification, and anomaly detectionStoring and using embeddings efficientlyGet started with Gemini Embedding 2Sort: