Exploring Text-Embedding-3-Large: A Comprehensive Guide to the new OpenAI Embeddings

OpenAI's latest text embedding models, text-embedding-3-large and text-embedding-3-small, are revolutionizing text analysis. This article explores their capabilities, applications, and practical usage.

Embeddings translate human language into machine-readable formats, crucial for AI tasks. OpenAI's new models significantly improve this process for developers and data scientists. We'll cover their core functions, applications, and effective implementation.

Understanding Text Embeddings

Text embeddings are numerical representations capturing the semantic meaning of text. They are essential for various NLP tasks, including sentiment analysis and text classification. Our guide, "Introduction to Text Embeddings with the OpenAI API," provides a comprehensive overview of using the OpenAI API for embedding creation.

Text Embeddings Illustration

Newcomers to embeddings should consult our "Introduction to Embeddings with the OpenAI API" course.

OpenAI's New Embedding Models

Released January 25, 2024, these models represent text in high-dimensional space for improved understanding. text-embedding-3-small prioritizes speed and storage, while text-embedding-3-large offers superior accuracy. The dimensions parameter allows adjusting text-embedding-3-large to 1536 dimensions (from its native 3072) without significant performance loss.

Benchmarking

text-embedding-3-large surpasses previous models (including text-embedding-ada-002) on MIRACL and MTEB benchmarks. The table below summarizes the comparison:

Model	Dimension	Max token	Knowledge cutoff	Pricing ($/1k tokens)	MIRACL average	MTEB average
ada v2	1536	8191	September 2021	0.0001	31.4	61.0
text-embedding-3-small				0.00002	44.0	62.3
text-embedding-3-large	3072			0.00013	54.9	64.6

Model

Dimension

Max token

Knowledge cutoff

Pricing ($/1k tokens)

MIRACL average

MTEB average

ada v2

1536

8191

September 2021

0.0001

31.4

61.0

text-embedding-3-small

0.00002

44.0

62.3

text-embedding-3-large

3072

0.00013

54.9

64.6

Higher dimensions in text-embedding-3-large (3072 vs. 1536) enhance performance but increase cost. Model selection depends on task requirements (multilingual needs, text complexity, budget). text-embedding-3-large excels in complex, multilingual scenarios, while text-embedding-3-small suits budget-conscious applications.

Applications

Both models find diverse applications:

text-embedding-3-large Applications:

Applications of text-embedding-3-large (images generated using GPT-4)

• Multilingual customer support automation (18 languages)
• Advanced semantic search engines
• Cross-lingual content recommendation systems

text-embedding-3-small Applications:

Applications of text-embedding-3-small (Image generated using GPT-4)

• Cost-effective sentiment analysis
• Scalable content categorization
• Efficient language learning tools

Step-by-Step Guide: Document Similarity

This guide uses the CORD-19 dataset (available on Kaggle) to demonstrate document similarity using all three models. Install necessary libraries:

pip -q install tiktoken openai

Import libraries:

import os
import tiktoken
import numpy as np
import pandas as pd
from openai import OpenAI
from sklearn.metrics.pairwise import cosine_similarity

Load and preprocess data (a 1000-document sample is used for brevity):

scientific_docs = pd.read_parquet("./data/cord19_df_sample.parquet")

def concatenate_columns_with_null_handling(df, body_text_column, abstract_column, title_column, new_col_name):
    df[new_col_name] = df[body_text_column].fillna('') + df[abstract_column].fillna('') + df[title_column].fillna('')
    return df

new_scientific_docs = concatenate_columns_with_null_handling(scientific_docs, "body_text", "abstract", "title", "concatenated_text")

def num_tokens_from_text(text: str, encoding_name="cl100k_base"):
    encoding = tiktoken.get_encoding(encoding_name)
    num_tokens = len(encoding.encode(text))
    return num_tokens

new_scientific_docs['num_tokens'] = new_scientific_docs["concatenated_text"].apply(lambda x: num_tokens_from_text(x))
smaller_tokens_docs = new_scientific_docs[new_scientific_docs['num_tokens'] 
<p>Set OpenAI API key and create client:</p>
<pre class="brush:php;toolbar:false">os.environ["OPENAI_API_KEY"] = "YOUR KEY"
client = OpenAI()

Generate embeddings:

def get_embedding(text_to_embbed, model_ID):
    text = text_to_embbed.replace("\n", " ")
    return client.embeddings.create(input=[text_to_embbed], model=model_ID).data[0].embedding

smaller_tokens_docs_reset['text-embedding-3-small'] = smaller_tokens_docs_reset["concatenated_text"].apply(lambda x: get_embedding(x, "text-embedding-3-small"))
smaller_tokens_docs_reset['text-embedding-3-large'] = smaller_tokens_docs_reset["concatenated_text"].apply(lambda x: get_embedding(x, "text-embedding-3-large"))
smaller_tokens_docs_reset['text-embedding-ada-002'] = smaller_tokens_docs_reset["concatenated_text"].apply(lambda x: get_embedding(x, "text-embedding-ada-002"))

Find similar documents using cosine similarity:

def find_top_N_similar_documents(df, chosen_index, embedding_column_name, top_N=3):
    chosen_document_embedding = np.array(df.iloc[chosen_index][embedding_column_name]).reshape(1, -1)
    embedding_matrix = np.vstack(df[embedding_column_name])
    similarity_scores = cosine_similarity(chosen_document_embedding, embedding_matrix)[0]
    df_temp = df.copy()
    df_temp['similarity_to_chosen'] = similarity_scores
    similar_documents = df_temp.drop(index=chosen_index).sort_values(by='similarity_to_chosen', ascending=False)
    top_N_similar = similar_documents.head(top_N)
    return top_N_similar[["concatenated_text", 'similarity_to_chosen']]

chosen_index = 0
top_3_similar_3_small = find_top_N_similar_documents(smaller_tokens_docs_reset, chosen_index, "text-embedding-3-small")
top_3_similar_3_large = find_top_N_similar_documents(smaller_tokens_docs_reset, chosen_index, "text-embedding-3-large")
top_3_similar_ada_002 = find_top_N_similar_documents(smaller_tokens_docs_reset, chosen_index, "text-embedding-ada-002")

print("Top 3 Similar Documents with:")
print("--> text-embedding-3-small")
print(top_3_similar_3_small)
print("\n")
print("--> text-embedding-3-large")
print(top_3_similar_3_large)
print("\n")
print("--> text-embedding-ada-002")
print(top_3_similar_ada_002)
print("\n")

Conclusion

OpenAI's new embedding models offer substantial improvements in NLP. The choice between text-embedding-3-large and text-embedding-3-small depends on the specific application's needs, balancing accuracy and cost. This guide provides the tools to effectively utilize these powerful models in various projects. Further resources on the OpenAI API and fine-tuning are available.

The above is the detailed content of Exploring Text-Embedding-3-Large: A Comprehensive Guide to the new OpenAI Embeddings. For more information, please follow other related articles on the PHP Chinese website!