Gemini Embedding 2: Google's New Multimodal AI Embedding System

1. What Is an Embedding?

An embedding is a numerical representation of a piece of data (text, image, audio, etc.) in the form of a high-dimensional vector. The core idea is that elements with similar meaning produce vectors that are close together in that space.

This property makes embeddings the foundation of technologies such as semantic search (finding documents by meaning, not exact words), RAG systems (Retrieval-Augmented Generation), and recommendation engines.

2. What Is Gemini Embedding 2?

Gemini Embedding 2 (gemini-embedding-2-preview) is Google's most recent embeddings model, available on Vertex AI since March 10, 2026. Its key differentiator compared to previous models like text-embedding-004 is that it is multimodal by design: it can generate embeddings from text, images, audio, video and PDFs, mapping all these types into a single shared vector space.

3. Key Advantages Over Previous Models

Google's previous embedding models (such as text-embedding-004 or multimodalembedding@001) had important limitations: they were unimodal or multimodal but with less context capacity and no task instructions. Gemini Embedding 2 surpasses these on several fronts.

// Task instruction examples
{
  "task_type": "RETRIEVAL_DOCUMENT",    // for indexing documents
  "task_type": "RETRIEVAL_QUERY",       // for search queries
  "task_type": "CODE_RETRIEVAL_QUERY",  // for code search
  "task_type": "SEMANTIC_SIMILARITY",  // for similarity comparison
  "task_type": "CLASSIFICATION",       // for classification
  "task_type": "CLUSTERING"            // for clustering
}

4. How to Use It: Practical Python Example

The model is accessed through Vertex AI. Here is a basic example of how to generate text and image embeddings in the same vector space:

import vertexai
from vertexai.preview.vision_models import MultiModalEmbeddingModel

vertexai.init(project="my-project", location="us-central1")

model = MultiModalEmbeddingModel.from_pretrained("gemini-embedding-2-preview")

# Text embedding with task instruction
text_embeddings = model.get_embeddings(
    contextual_text="How do microservices communicate in a distributed system?",
    task_type="RETRIEVAL_QUERY",
    output_dimensionality=1024  # reduce with MRL
)

print(f"Dimensions: {len(text_embeddings.text_embedding)}")
# Output: Dimensions: 1024

from vertexai.preview.vision_models import Image

# Image embedding
image = Image.load_from_file("architecture-diagram.png")

image_embeddings = model.get_embeddings(
    image=image,
    output_dimensionality=1024
)

# Text embedding (same query)
text_embeddings = model.get_embeddings(
    contextual_text="microservices architecture diagram",
    task_type="RETRIEVAL_QUERY",
    output_dimensionality=1024
)

# Now you can compare text with image directly
import numpy as np

def cosine_similarity(a, b):
    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

similarity = cosine_similarity(
    text_embeddings.text_embedding,
    image_embeddings.image_embedding
)
print(f"Text-image similarity: {similarity:.4f}")
# If high → the image is relevant for the text query

from vertexai.preview.vision_models import Video, VideoSegmentConfig

# Embedding a scanned PDF (automatic OCR)
pdf_embeddings = model.get_embeddings(
    video=None,
    image=None,
    contextual_text=None,
    pdf="gs://my-bucket/technical-report.pdf",  # Google Cloud Storage
    output_dimensionality=1024
)

# The model applies OCR internally — no preprocessing needed
print(f"Generated embedding: {len(pdf_embeddings.text_embedding)} dims")

5. Most Relevant Use Cases

6. Limitations to Keep in Mind

Conclusion

Gemini Embedding 2 is not an incremental update — it is a paradigm shift in how we think about embeddings. The ability to unify text, images, audio, video and PDFs in the same vector space opens up use cases that previously required multiple specialized models and complex fusion layers.

For those of us building RAG systems or semantic search, the task instructions and adjustable dimensions with MRL are features that in practice improve both accuracy and operational cost. It is worth evaluating now.