Gemini Embedding 2: One Vector Space for Five Modalities

Melvin Vivas · X post · 2026-04-23 · Open on X

Topics: Embeddings & Vector Databases, Retrieval-Augmented Generation (RAG) · Level: intermediate

Summary

Gemini Embedding 2 is now generally available. It is a single embedding model that maps text, images, video, audio and PDFs into one shared vector space, which allows search across modalities.

Key points

• Gemini Embedding 2 is generally available (GA).
• It embeds 5 modalities (text, images, video, audio, PDFs) into one unified space.
• It supports up to 8,192 input tokens and 100+ languages.
• It embeds audio natively, with no transcription step.
• The output size is flexible.

Resources mentioned

• Gemini Embedding 2 · tool · blog.google · check price · open in a browser to verify
  Google's multimodal embedding model, with one vector space for text, image, video, audio and PDF.

Try this

• Build a multimodal search/RAG system that searches text, images, audio and PDFs in one Gemini Embedding 2 index.