Where can I find the removed distance and nearest neighbor functions in the OpenAI Python library?

The latest version of the OpenAI Python library no longer includes the embeddings_utils module, which means the following import now fails:

PYTHONCopied
from openai.embeddings_utils import (get_embedding, distances_from_embeddings, indices_of_nearest_neighbors_from_distances)

I'm specifically looking for the implementations of:

• distances_from_embeddings
• indices_of_nearest_neighbors_from_distances

According to recent recommendations, these helper functions are no longer shipped with the library. Instead, it is advised to compute similarities directly (e.g., using cosine similarity) or to use a vector database for efficient nearest neighbor lookups. For example, since the OpenAI embeddings are normalized to length 1, you can compute cosine similarity using a simple dot product with NumPy. Here’s a basic replacement implementation:

PYTHONCopied
import numpy as np

def cosine_similarity(a, b):
    # Since a and b are already normalized, np.dot is sufficient.
    return np.dot(a, b)

def distances_from_embeddings(embeddings, query_embedding):
    # Compute cosine similarity (distance) between query_embedding and each embedding.
    return [cosine_similarity(query_embedding, emb) for emb in embeddings]

def indices_of_nearest_neighbors_from_distances(distances, k=5):
    # Get indices sorted by similarity (high to low)
    return sorted(range(len(distances)), key=lambda i: distances[i], reverse=True)[:k]

# Example usage:
embeddings = [np.random.rand(768) for _ in range(100)]  # dummy data; assume embeddings are normalized
query_embedding = np.random.rand(768)
query_embedding /= np.linalg.norm(query_embedding)  # normalize

similarities = distances_from_embeddings(embeddings, query_embedding)
top_indices = indices_of_nearest_neighbors_from_distances(similarities, k=3)
print(top_indices)

This basic implementation serves as a replacement for the removed helper functions. For scaling up or faster searches, consider using a dedicated vector database or libraries like FAISS.