Gemini Embedding 2 is Google's embedding generation model that's ideal for complex retrieval and analytics tasks.
Gemini Embedding 2 accepts multimodal inputs to generate 3072-dimensional vectors. It accepts images, text, documents, audio, and video inputs and semantically maps the generated vectors into a unified semantic space. This lets you perform tasks, such as searching for an image based on a text description.
Gemini Embedding 2 introduces several features to optimize embedding quality and flexibility:
-
Custom task instructions:By specifying task instructions—for example,
task:code retrievalortask:search result—optimize the embeddings for the intended relationships and retrieve more accurate results for the specific goal. -
Adjustable result size:The model generates a 3072-dimensional float vector, by default. However, you can retrieve a smaller dimensional output by specifying the
output_dimensionalityparameter. -
Document OCR:Read OCR from document inputs.
-
Audio track extraction:Extract audio tracks from video inputs and interleave them with video frames.
Try in Vertex AI (Preview) Deploy example app
gemini-embedding-2-preview
- Inputs:
Text , Images , Audio , Video , PDF - Outputs:
Embeddings
- Maximum input tokens: 8,192
- Maximum output tokens: N/A
8,192 tokens
Up to 3,072 (with MRL support)
- Supported
- Not supported
- Maximum images per prompt: 6
- Maximum file size per file for inline data or direct uploads through the console: No limit
- Maximum file size per file from Google Cloud Storage: No limit
- Maximum number of output images per prompt: N/A
- Supported MIME types:
image/png,image/jpeg
- Maximum number of files per prompt: 1
- Maximum number of pages per file: 6
- Maximum file size per file: N/A
- Supported MIME types:
application/pdf
- Maximum video length (with audio): 80 seconds
- Maximum video length (without audio): 120 seconds
- Maximum number of videos per prompt: 1
- Supported MIME types:
video/mpeg,video/mp4
- Maximum audio length per prompt: 80 seconds
- Maximum number of audio files per prompt: 1
- Supported MIME types:
audio/mp3,audio/wav
Model availability
- United States
- us-central1
-
gemini-embedding-2-preview - Launch stage: Public preview
- Release date: March 10, 2026
