Vick's

Have you ever wondered how Google, ChatGPT, or smart assistants find relevant answers to your questions? The secret behind this is something called vector embeddings and similarity search!

Today, we'll break this down step by step and even visualize it in 3D—so it feels like you're seeing AI thinking in real time! 🤯

🤔 What’s the Problem?

Imagine you have a book filled with random sentences. If someone asks, "What is AI?", how would you find the most relevant answer from the book?

One way is to read the whole book. But that's too slow! Instead, computers use a special trick:

1️⃣ Convert words into numbers (because computers love numbers).
2️⃣ Store these numbers in memory (so they can quickly search).
3️⃣ Find the closest match when someone asks a question.

This is called vector search, and it helps AI quickly find the most relevant answers!

💡 How Do Computers Understand Words?

Computers don’t understand words the way we do. Instead, they convert words into vectors (a fancy term for a list of numbers).

For example:

Notice how similar sentences have similar numbers? That’s because AI understands meaning through math!

🔍 Step-by-Step: How We Find Answers

Let’s say we have five sentences and want to find which one is most related to "What is AI?". We’ll use:

• Sentence Transformers (to convert words into numbers).
• FAISS (Facebook AI Similarity Search) (to find the closest match).
• PCA (Principal Component Analysis) (to reduce dimensions and make it visual).

🚀 Step 1: Convert Sentences into Numbers

We take a few sentences and convert them into vector embeddings.

from sentence_transformers import SentenceTransformer

# Load an AI model to generate vectors
model = SentenceTransformer("all-MiniLM-L6-v2")

# Sample sentences
documents = [
    "AI is transforming the world.",
    "Machine learning is a subset of AI.",
    "Vector databases help store embeddings.",
    "FAISS is great for similarity search.",
    "Deep learning is part of AI."
]

# Convert text to vector embeddings
embeddings = model.encode(documents)

🎯 What Happens Here?

• Each sentence becomes a vector (a list of numbers).
• Similar sentences have similar numbers (so AI can match them easily).

⚡ Step 2: Store These Vectors in FAISS (AI’s Fast Memory!)

FAISS is like a super-smart dictionary that stores word meanings in number form and helps AI find them super fast!

import faiss
import numpy as np

# Convert embeddings to float format
embeddings = np.array(embeddings).astype('float32')

# Initialize FAISS index (it helps AI quickly compare sentences)
dimension = embeddings.shape[1]
index = faiss.IndexFlatL2(dimension)

# Store the vectors inside FAISS
index.add(embeddings)

🎯 What Happens Here?

• FAISS acts like a superfast brain that helps AI find similar words instantly!

🎯 Step 3: Find the Most Similar Answer to Our Question

Let’s say we ask "What is AI?". FAISS will compare it to all stored sentences and return the closest match!

# Convert our question into a vector
query = "What is AI?"
query_embedding = model.encode([query]).astype('float32')

# Search for the most similar sentence
k = 2  # Number of closest matches we want
distances, indices = index.search(query_embedding, k)

# Show the most relevant answer
print("Query:", query)
for i in range(k):
    print(f"Match {i+1}: {documents[indices[0][i]]} (Distance: {distances[0][i]:.4f})")

🎯 What Happens Here?

• FAISS compares our question to stored sentences.
• It finds the closest matches using mathematical distance.
• The smaller the distance, the more similar the answer!

🎨 Step 4: Visualizing AI's Thought Process in 3D!

Since vectors have hundreds of dimensions, we reduce them to 3D and plot them.

import matplotlib.pyplot as plt
from sklearn.decomposition import PCA
from mpl_toolkits.mplot3d import Axes3D

# Reduce 384D embeddings to 3D for visualization
pca = PCA(n_components=3)
reduced_embeddings = pca.fit_transform(embeddings)

# 3D Scatter Plot
fig = plt.figure(figsize=(10, 7))
ax = fig.add_subplot(111, projection='3d')

# Plot all document points
for i, (x, y, z) in enumerate(reduced_embeddings):
    ax.scatter(x, y, z, label=f"Doc {i+1}")
    ax.text(x, y, z, f"{i+1}", fontsize=12)

# Draw similarity lines from the first query to closest matches
query_point = reduced_embeddings[0]
for i in range(1, k):  # Ignore self-match
    match_point = reduced_embeddings[indices[0][i]]
    ax.plot([query_point[0], match_point[0]], 
            [query_point[1], match_point[1]], 
            [query_point[2], match_point[2]], 'r--')

# Labels and title
ax.set_title("3D Visualization of Document Embeddings")
ax.set_xlabel("PCA 1")
ax.set_ylabel("PCA 2")
ax.set_zlabel("PCA 3")
plt.legend()
plt.show()

🎯 What Happens Here?

• Each sentence becomes a point in 3D space.
• AI connects the most relevant answers with red dashed lines.
• You can literally see AI thinking!

🎯 Conclusion: How AI Thinks in Numbers

✅ AI converts text into numbers using embeddings.
✅ Similar sentences have similar numbers (so AI can match them).
✅ FAISS quickly searches for the closest match using math.
✅ We can visualize AI’s thought process in 3D!