Semantic Search Using Text Embeddings (With ChatGPT + Python)

Semantic Search Using Text Embeddings (With ChatGPT + Python)

Traditional keyword-based search is outdated. In today’s world, users expect search engines to understand the meaning behind their queries — not just the words.

That’s where semantic search comes in. It’s a technique that uses AI-generated embeddings to understand and compare the meanings of words, phrases, or even entire documents.

🔍 Why Semantic Search?

• Finds similar content even when keywords differ

• Understands natural language queries

• Great for product search, recommendation systems, document retrieval, and more

🛠 Tools Used

• ChatGPT / OpenAI API – for generating embeddings

• Python – our programming language

• Pandas – for data manipulation

• NumPy – for vector math

• tiktoken – to estimate token usage and cost

📊 Step-by-Step Guide to Semantic Search

1. Generate Embeddings
Each word or phrase is converted into a high-dimensional vector using an OpenAI model:

df['embedding'] = df['Words'].apply(lambda x: get_embedding(x))

2. Define Cosine Similarity Function
This function compares vectors to measure how similar their meanings are:

import numpy as np

def cosine_similarity(vec1, vec2):
    vec1 = np.array(vec1).flatten()
    vec2 = np.array(vec2).flatten()

    if vec1.shape != vec2.shape:
        raise ValueError("Vectors must be of same dimension")

    dot_product = np.dot(vec1, vec2)
    norm_a = np.linalg.norm(vec1)
    norm_b = np.linalg.norm(vec2)

    if norm_a == 0 or norm_b == 0:
        raise ValueError("Cosine similarity is not defined for zero vectors")

    return dot_product / (norm_a * norm_b)

3. Combine Concepts
Want to find words similar to both “Sunflower” and “Lotus”?

v1 = df['embedding'].iloc[7]  # Sunflower
v2 = df['embedding'].iloc[1]  # Lotus
v = v1 + v2

df['similarities'] = df['embedding'].apply(lambda x: cosine_similarity(x, v))
top_matches = df.sort_values('similarities', ascending=False).head(10)

4. Semantic Search from User Query

search_term = "yellow flower"
search_term_vector = get_embedding(search_term)

df['similarities'] = df['embedding'].apply(lambda x: cosine_similarity(x, search_term_vector))
results = df.sort_values('similarities', ascending=False).head(10)

5. Estimate Token Cost

import tiktoken

words = list(df['Words'])
enc = tiktoken.encoding_for_model('text-embedding-3-small')
total_tokens = sum(len(enc.encode(word)) for word in words)

cost_per_token = 0.02 / 1_000_000
estimated_cost = total_tokens * cost_per_token

print(f"Total tokens: {total_tokens}")
print(f"Estimated cost: ${estimated_cost:.10f}")

🧭 Visual Overview

Imagine each word as a dot in space. Embeddings place similar concepts closer together.
When we calculate cosine similarity, we are measuring how “aligned” these meaning-vectors are.

You can upload a diagram to illustrate this or let me know — I can give you one.

🚀 Conclusion

Semantic search is a game-changer for modern applications. Instead of matching exact words, you’re matching meaning — thanks to the power of embeddings and vector math.

Whether you're building an internal search engine, recommendation system, or a chatbot — this technique can dramatically improve user experience.

💡 Pro Tips

• Use text-embedding-3-large for better accuracy

• Use vector databases like Pinecone, FAISS, or Weaviate for scale

• Normalize vectors if needed for better cosine distance performance