Gemini Embedding 2 reached general availability

Google made Gemini Embedding 2 generally available on April 22. The release announcement describes a natively multimodal embedding model intended to represent text, images, video and audio for search and reasoning systems. Google said the GA release adds the stability and optimizations needed to move projects beyond preview, with access through the Gemini API and Gemini Enterprise Agent Platform.

Multimodal retrieval changes the index contract

A shared embedding space can reduce the number of modality-specific retrieval pipelines a team operates. It does not make their inputs interchangeable. Video segmentation, audio channel selection, image preprocessing and text chunking still determine what an embedding represents. Moving to a new model can also change nearest neighbors even when the source corpus is untouched.

That makes the embedding model and preprocessing configuration part of the index version. Store both alongside vectors, and keep enough source metadata to explain why an item was returned. A GA label is a lifecycle signal from Google; it is not evidence that an existing relevance threshold will transfer unchanged.

Plan migration as a measured re-index

Teams adopting the model should build an evaluation set from real queries across every modality they plan to support. Compare retrieval quality, latency, index size and failure handling against the current system. Include cross-modal queries, such as text seeking a moment in video, rather than evaluating each medium in isolation.

For migration, create a parallel index and shadow representative traffic before switching reads. Define rollback around index versions, not just an API configuration toggle. Re-embedding a large corpus can be the dominant cost and time constraint, so checkpoint progress and verify coverage before cutover.