Gemini Embedding 2 Preview: First Natively Multimodal Embedding Model

​

Key Highlights

• First Natively Multimodal Embedding: Unified vector space for text, image, video, audio, and PDF
• MTEB English #1: 68.32 score, leading in classification (+9.6), retrieval (+9.0), and clustering (+3.7)
• Flexible Dimensions: Default 3072, supports 128–3072 truncation via Matryoshka Representation Learning (MRL); 768-dim still achieves 67.99
• Extended Input: Up to 8192 text tokens, 6 images/request, 120-second video
• 100+ Languages: Multilingual embeddings, MTEB Multilingual Top 5

​

Background

​

Detailed Analysis

​

Core Features

​

Performance Highlights

​

Comparison with Previous Models

​

Multimodal Input Specifications

​

Supported Task Types

​

Technical Specifications

​

Practical Applications

​

Recommended Use Cases

1. Cross-Modal Semantic Search: Search images with text, retrieve documents with images — unified vector space enables mixed retrieval
2. Multilingual RAG: 100+ languages for building global retrieval-augmented generation systems
3. Document Intelligence: Embed PDFs directly without preprocessing for document retrieval
4. Video/Audio Content Retrieval: Native video and audio embedding for media content management
5. Clustering & Classification: +9.6 classification and +3.7 clustering advantage for large-scale content organization
6. Code Semantic Search: Query code snippets with natural language to boost developer productivity

​

Code Examples

​

Text Embedding

from openai import OpenAI

client = OpenAI(
    api_key="your-apiyi-key",
    base_url="https://api.apiyi.com/v1"
)

response = client.embeddings.create(
    model="gemini-embedding-2-preview",
    input="What are the key features of Google's latest multimodal embedding model?",
    dimensions=768  # Optional: 128–3072
)

embedding = response.data[0].embedding
print(f"Dimensions: {len(embedding)}")

​

Batch Text Embedding

texts = [
    "Latest trends in artificial intelligence",
    "Machine learning applications in healthcare",
    "How large language models work"
]

response = client.embeddings.create(
    model="gemini-embedding-2-preview",
    input=texts,
    dimensions=1536
)

for i, data in enumerate(response.data):
    print(f"Text {i}: {len(data.embedding)} dimensions")

​

Semantic Search Example

import numpy as np

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

# Build document embeddings
docs = ["Quantum computing principles", "Deep learning intro", "Blockchain overview"]
doc_resp = client.embeddings.create(
    model="gemini-embedding-2-preview",
    input=docs,
    dimensions=768
)
doc_embeddings = [d.embedding for d in doc_resp.data]

# Query
query_resp = client.embeddings.create(
    model="gemini-embedding-2-preview",
    input="How do neural networks work?",
    dimensions=768
)
query_embedding = query_resp.data[0].embedding

# Calculate similarity
for i, doc_emb in enumerate(doc_embeddings):
    sim = cosine_similarity(query_embedding, doc_emb)
    print(f"{docs[i]}: {sim:.4f}")

​

Best Practices

1. Choose the right dimensions: 768-dim offers the best value (67.99 score, half storage), 3072-dim for maximum precision
2. Normalize after truncation: 3072-dim vectors are pre-normalized; smaller dimensions require manual normalization
3. Use prompt instructions: Differentiate query vs. document side for retrieval to significantly improve results
4. Don’t mix versions: Incompatible with text-embedding-004 or gemini-embedding-001 vectors — migration requires full rebuild

​

Pricing and Availability

​

Pricing

​

Price Comparison

​

Deposit Bonus

​

Available Models

​

How to Access

• Website: apiyi.com
• API Endpoint: https://api.apiyi.com/v1
• Interface: /v1/embeddings (OpenAI-compatible)
• Works with all OpenAI SDKs

​

Summary and Recommendations

• Multimodal Unity: Text/image/video/audio/PDF share one vector space — one model for all retrieval
• Performance Leader: MTEB 68.32 #1, major leads in classification, retrieval, and clustering
• Flexible Dimensions: MRL supports 128–3072, balance precision vs. cost as needed
• Extended Input: 8192 tokens, 4x the previous generation

1. Cross-modal retrieval: Gemini Embedding 2 is the only choice — and the best one
2. Text-only + lowest cost: text-embedding-3-small remains the cheapest option
3. Text-only + high accuracy: Gemini Embedding 2 at 768-dim already surpasses text-embedding-3-large
4. RAG scenarios: 8192 token input + flexible dimensions, ideal for large document chunking and retrieval

• Applications requiring cross-modal search (image-to-text, text-to-image, etc.)
• Developers building multilingual RAG systems
• Enterprise scenarios processing PDF/video/audio content
• Retrieval systems seeking the highest embedding quality