Moving from OpenAI Embeddings to Cohere — Quality Comparison

I've been using OpenAI's embeddings for a while in my NLP project, but I've recently started experimenting with Cohere's embeddings due to some specific needs in my application. I'm curious about how others have experienced the quality difference between the two.

I ran a few tests comparing the two models on semantic similarity tasks. Using a small dataset of about 1,000 sentence pairs, I calculated cosine similarity scores. With OpenAI's embeddings, I was getting an average score of 0.85 on the similarity metric, whereas Cohere seemed to yield around 0.82 on the same pairs. While this isn't a huge drop, I did notice that Cohere sometimes captured more nuanced meanings, especially with domain-specific vocabulary.

Here's the Python code snippet I used for comparison:

import numpy as np
from openai import Embedding
from cohere import Client

# Assuming you have initialized the OpenAI API and Cohere Client
openai_embeddings = [Embedding.create(input=sentence)['data'][0]['embedding'] for sentence in sentences]
cohere_client = Client('YOUR_COHERE_API_KEY')
cohere_embeddings = [cohere_client.embed(text=sentence).embeddings[0] for sentence in sentences]

# Calculate cosine similarity function
def cosine_similarity(a, b):
    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

# Getting similarity scores
similarity_scores = [cosine_similarity(openai_embeddings[i], cohere_embeddings[i]) for i in range(len(sentences))]

I’m still fine-tuning how I’m using these embeddings in downstream tasks, but I'd love to hear your insights. Has anyone else gone through this transition? What are your thoughts on quality, cost, or ease of integration between OpenAI and Cohere?