Cohere Embed v4: The First Production Multimodal Embedding

Cohere Revolutionizes Embeddings with Multimodal

Cohere has announced the general availability of Embed v4 Multimodal, a major advancement in the world of embeddings. For the first time, a production model can vectorize text, images, and mixed documents (PDFs, slides, tables) into the same semantic space.

"Embed v4 eliminates the complexity of document parsing," declares Aidan Gomez, CEO of Cohere. "You can now vectorize a PDF as-is, with its images, tables, and text, without preprocessing."

Benchmark Performance

MTEB Results

The Real Innovation: Multimodal

The MTEB score doesn't tell the whole story. Embed v4 excels where others don't exist:

*Only a few experimental models support images

To understand the importance of multimodal in embeddings, check out our guide on multimodal RAG.

Technical Innovations

Unified Text-Image Embedding

Embed v4 creates a vector space where text and images coexist:

DEVELOPERpython
import cohere

co = cohere.ClientV2('your-api-key')

# Text embedding
text_response = co.embed(
    texts=["Product description"],
    model="embed-v4",
    input_type="search_document",
    embedding_types=["float"]
)

# Image embedding (base64 or URL)
image_response = co.embed(
    images=["data:image/jpeg;base64,..."],
    model="embed-v4",
    input_type="image",
    embedding_types=["float"]
)

# Both embeddings are in the same semantic space!

Technical Specifications

Matryoshka Embeddings

Embed v4 supports Matryoshka embeddings, allowing dimension reduction without re-encoding:

DEVELOPERpython
# Full dimensions (1536)
full_embedding = co.embed(
    texts=["Your text"],
    model="embed-v4",
    embedding_types=["float"]
)

# Reduced dimensions (256) - same vector truncated
compact_embedding = co.embed(
    texts=["Your text"],
    model="embed-v4",
    embedding_types=["float"],
    output_dimension=256  # Matryoshka truncation
)

This approach optimizes the cost/quality tradeoff without regenerating all your embeddings.

Impact on RAG Pipelines

End of Complex Parsing

Before Embed v4, vectorizing a PDF required:

1. Text extraction (PyPDF, pdfplumber)
2. Image OCR (Tesseract, Azure Vision)
3. Table detection (Camelot, Tabula)
4. Context reconstruction
5. Separate chunking and embedding

With Embed v4:

1. Screenshot or image of PDF
2. Direct embedding

"We removed 80% of our preprocessing pipeline," testifies Marie Laurent, CTO of a French legaltech startup. "Retrieval quality improved because the model sees documents like a human does."

Transformed Use Cases

Visual E-commerce

• Product image search
• PDF catalogs vectorized as-is
• Technical sheets with diagrams

Technical Documentation

• Manuals with diagrams
• Architecture schemas
• Annotated screenshots

Legal and Finance

• Scanned contracts
• Reports with charts
• Filled forms

Check out our guide on e-commerce RAG for concrete examples.

Pricing and Availability

Pricing

Comparison with Competition

Availability

Embed v4 is available on:

• Direct Cohere API
• Amazon Bedrock
• Amazon SageMaker JumpStart
• Azure AI Foundry
• Google Cloud Vertex AI

Practical Integration

Complete Example: Multimodal RAG

DEVELOPERpython
import cohere
from qdrant_client import QdrantClient

co = cohere.ClientV2('your-api-key')
qdrant = QdrantClient(url="http://localhost:6333")

# Index a PDF as image
def index_pdf_page(image_base64, metadata):
    response = co.embed(
        images=[f"data:image/png;base64,{image_base64}"],
        model="embed-v4",
        input_type="image",
        embedding_types=["float"]
    )

    qdrant.upsert(
        collection_name="documents",
        points=[{
            "id": metadata["id"],
            "vector": response.embeddings.float[0],
            "payload": metadata
        }]
    )

# Search by text (cross-modal)
def search_by_text(query):
    query_embedding = co.embed(
        texts=[query],
        model="embed-v4",
        input_type="search_query",
        embedding_types=["float"]
    )

    # Find relevant images/PDFs with a text query
    results = qdrant.search(
        collection_name="documents",
        query_vector=query_embedding.embeddings.float[0],
        limit=5
    )
    return results

Best Practices

1. Choose the Right input_type

• search_document: Text to index
• search_query: User query
• image: Images to index or search

2. Optimize Images

• Ideal resolution: 1024x1024 pixels
• Maximum: 2 megapixels
• Formats: JPEG for photos, PNG for captures

3. Batching

DEVELOPERpython
# Up to 96 texts or 1000 images per request
response = co.embed(
    images=list_of_images[:1000],
    model="embed-v4",
    input_type="image"
)

Our Take

Embed v4 Multimodal is a decisive advancement for RAG applications handling rich documents. The ability to vectorize PDFs, presentations, and images without complex preprocessing radically simplifies architectures.

Strengths:

• First production multimodal
• 128K token context
• Matryoshka for cost optimization
• Native cloud integration

Points to watch:

• Pure text MTEB score lower than Qwen3/Gemini
• Higher price for large image volumes

For new projects with visual documents, Embed v4 is our recommendation. For pure text at very high volume, consider Qwen3-Embedding (open source) or Google Gemini Embedding.

Explore our comprehensive guide on choosing embedding models to deepen this decision.