Gemini Embeddings 2: text, image, video, audio in one vector space

My notes

Summary

Google released Gemini Embeddings 2, the first model that natively maps text, images, video, audio, and documents into a single vector space without transcription or conversion. This eliminates the need for separate pipelines per modality and reduces information loss, latency, and retrieval errors compared to existing multimodal approaches.

Key Insight

• Native multimodal understanding, not conversion: Unlike other “multimodal” embedding models that secretly transcribe audio to text or describe video frames, Gemini Embeddings 2 processes each modality directly. This preserves tone, visual detail, and context that gets lost in conversion.
• Benchmark leader: Outperforms Amazon Nova, Voyage multimodal models, and Google’s own previous embedding models across all data types.
• Single vector space: Text, images, video, audio, and documents all map to the same embedding space. A text query can retrieve a relevant video clip or audio segment directly, no stitching of five different retrieval pipelines.
• Practical impact for RAG: Teams with heterogeneous knowledge bases (meeting recordings, product images, support call audio, documents) can build a single retrieval pipeline instead of maintaining separate transcription, image description, and embedding workflows.
• Available now: Public preview via the Gemini API.