text-embedding-004

Google's latest Gecko-based embedding model featuring a 768-dimensional output and a 2,048-token context window for state-of-the-art retrieval.

Text-embedding-004 (Gecko) represents a significant leap in semantic representation, outperforming previous iterations on the MTEB benchmark with a 54.5 score. It processes up to 2,048 tokens per request and supports Matryoshka Representation Learning (MRL), allowing developers to truncate embeddings to smaller sizes (like 128 or 256) without losing significant accuracy. This model is optimized for RAG pipelines, semantic search, and classification tasks via the Gemini API, providing a cost-effective solution for high-density vector search.

https://ai.google.dev/gemini-api/docs/embeddings