cohere

Exploring Cohere's Embed-English-Light-V3.0: Efficient Text Embedding for Modern Applications

Tal Peretz

— 2 min read
Exploring Cohere's Embed-English-Light-V3.0: Efficient Text Embedding for Modern Applications

The Cohere Embed-English-Light-V3.0 model is a state-of-the-art addition to Cohere's lineup, providing efficient and powerful text embeddings at a competitive price. This lightweight model, with its 384 dimensions, offers performance close to the full-sized Cohere Embed-English-V3.0 while being more resource-efficient.

Dimensions and Performance

With 384 dimensions, this model is designed to maintain robust performance metrics, despite being smaller than its 1024-dimension counterpart. It is tuned for fast and effective text processing, scoring 62.0 on MTEB and 52.0 on BEIR benchmarks, slightly below the full-size version but impressive given its compactness.

Usage and Integration

Integration with the Cohere Embed-English-Light-V3.0 is seamless through the Cohere API, AWS SageMaker, AWS Bedrock, or private deployments. Developers can leverage the Cohere or AWS SDKs to embed this model into their applications. When using the model, specify input_type="search_document" for passage embedding and input_type="search_query" for query embedding.

Technical Capabilities

  • Max Input Tokens: Supports up to 512 input tokens.
  • Similarity Metric: Uses cosine or dot product similarity metrics for accurate results.

Applications

This model excels in applications such as semantic search, retrieval augmented generation (RAG), text classification, and document clustering. It is particularly effective for scenarios where short queries need to retrieve medium-length text passages.

Deployment Options

Deployment is flexible with options like the Cohere API, AWS Bedrock, and AWS SageMaker. Additionally, private deployments on your own hardware are available by contacting Cohere's sales team. Installation of the Cohere SDK is straightforward with pip install -U cohere, requiring an API key for access.

Embeddings created by this model can be stored in vector databases such as Zilliz Cloud or Pinecone, facilitating efficient similarity searches. Below is a practical example of using the Cohere API to generate and search embeddings:

import cohere
import numpy as np

cohere_key = "{YOUR_COHERE_API_KEY}"
co = cohere.Client(cohere_key)

docs = ["The capital of France is Paris", "PyTorch is a machine learning framework based on the Torch library.", "The average cat lifespan is between 13-17 years"]
doc_emb = co.embed(docs, input_type="search_document", model="embed-english-light-v3.0").embeddings

query = "What is Pytorch"
query_emb = co.embed([query], input_type="search_query", model="embed-english-light-v3.0").embeddings

scores = np.dot(np.asarray(query_emb), np.asarray(doc_emb).T)
max_idx = np.argsort(-scores)

print(f"Query: {query}")
for idx in max_idx:
    print(f"Score: {scores[idx]:.2f}")
    print(docs[idx])
    print("--------")

This code snippet demonstrates the process of embedding documents and queries, calculating similarity scores, and identifying the most relevant documents based on a query.

Read more