Cohere Embeddings

Cohere Embeddings Reference

Official Resources

• Docs & Cookbooks: https://github.com/cohere-ai/cohere-developer-experience
• API Reference: https://docs.cohere.com/reference/about

Models Overview

Input Types (CRITICAL)

Using the wrong input_type will silently degrade search quality. Cohere uses asymmetric embeddings where documents and queries are embedded differently.

Example: Search Pipeline

import cohere
co = cohere.ClientV2()

# INDEXING: Use search_document for docs you're storing
doc_response = co.embed(
    model="embed-english-v3.0",
    texts=documents,
    input_type="search_document"  # MUST use for storage
)

# QUERYING: Use search_query for user queries
query_response = co.embed(
    model="embed-english-v3.0",
    texts=[user_query],
    input_type="search_query"  # MUST use for retrieval
)

Native SDK Embeddings

Basic Text Embedding

response = co.embed(
    model="embed-english-v3.0",
    texts=["Hello world", "Machine learning is cool"],
    input_type="search_document"
)

embeddings = response.embeddings.float_
print(f"Embedding shape: {len(embeddings)} x {len(embeddings[0])}")

Embed v4 with Matryoshka Dimensions

# High precision (default)
response = co.embed(
    model="embed-v4.0",
    texts=["text"],
    input_type="search_document",
    output_dimension=1536
)

# Balanced (3x faster search)
response = co.embed(
    model="embed-v4.0",
    texts=["text"],
    input_type="search_document",
    output_dimension=512
)

# Compact (6x faster search)
response = co.embed(
    model="embed-v4.0",
    texts=["text"],
    input_type="search_document",
    output_dimension=256
)

Different Embedding Types

response = co.embed(
    model="embed-english-v3.0",
    texts=["Hello"],
    input_type="search_document",
    embedding_types=["float", "int8", "uint8", "binary", "ubinary"]
)

float_emb = response.embeddings.float_
int8_emb = response.embeddings.int8
binary_emb = response.embeddings.binary

Multimodal Embeddings (Embed v4)

Image Embeddings

import base64

with open("image.jpg", "rb") as f:
    image_base64 = base64.b64encode(f.read()).decode()

image_uri = f"data:image/jpeg;base64,{image_base64}"

response = co.embed(
    model="embed-v4.0",
    images=[image_uri],
    input_type="image"
)

Mixed Content

response = co.embed(
    model="embed-v4.0",
    inputs=[
        {"text": "A description of the product"},
        {"image": image_uri},
        {"text": "Another text chunk"}
    ],
    input_type="search_document"
)

Batch Processing

Hard Limit: 96 Items Per Request

def embed_in_batches(texts: list, batch_size: int = 96):
    """Embed texts in batches of 96 (Cohere API limit)."""
    all_embeddings = []

    for i in range(0, len(texts), batch_size):
        batch = texts[i:i + batch_size]
        response = co.embed(
            model="embed-english-v3.0",
            texts=batch,
            input_type="search_document"
        )
        all_embeddings.extend(response.embeddings.float_)

    return all_embeddings

Embed Jobs API (Large Datasets)

job = co.embed_jobs.create(
    model="embed-english-v3.0",
    dataset_id="your-dataset-id",
    input_type="search_document"
)

status = co.embed_jobs.get(job.job_id)
print(status.status)  # "processing", "complete", "failed"

LangChain Integration

Basic Usage

from langchain_cohere import CohereEmbeddings

embeddings = CohereEmbeddings(model="embed-english-v3.0")

vector = embeddings.embed_query("What is machine learning?")
vectors = embeddings.embed_documents(["Document 1", "Document 2"])

With Vector Store

from langchain_cohere import CohereEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_text_splitters import RecursiveCharacterTextSplitter

embeddings = CohereEmbeddings(model="embed-english-v3.0")

text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
docs = text_splitter.split_documents(your_documents)

vectorstore = FAISS.from_documents(docs, embeddings)
results = vectorstore.similarity_search("your query", k=5)

Best Practices

1. Match input types: Always use search_document for stored docs and search_query for queries
2. Batch efficiently: Hard limit of 96 texts per request
3. Choose dimensions wisely: Lower dimensions = faster search but slightly less precision
4. Chunk long texts: Consider chunking at ~6000 chars (texts auto-truncate at 8K)

from langchain_text_splitters import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(
    chunk_size=6000,
    chunk_overlap=200
)
chunks = splitter.split_text(long_document)