Author(s): Kapil Deshmukh

Originally published on Towards AI.

From embeddings to retrieval: mastering vector similarity search for next-gen enterprise AI applications.

🚀 Introduction

If you’ve ever asked ChatGPT a question like “What animal likes milk?” and received a precise answer, you’ve already seen vector similarity search in action.

Behind the scenes, AI systems rely on embeddings (numerical representations of text) and vector databases to find the most relevant information. This is the backbone of Retrieval-Augmented Generation (RAG), agentic AI, and modern enterprise AI chatbots.

In this article, we’ll explore:

• How embeddings transform text into vectors.
• The role of similarity metrics (cosine, Euclidean, dot product).
• How vector databases like FAISS, Pinecone, and Weaviate enable lightning-fast search.
• A hands-on Python demo with OpenAI embeddings.

🧠 What Are Embeddings?

Think of embeddings as the language of meaning for AI.

• Text is converted into high-dimensional vectors (e.g., 1536 dimensions for OpenAI).
• Similar concepts produce vectors that are close together.

Example:

"cat" → [0.8, 0.1, 0.2, …]

"dog" → [0.7, 0.15, 0.25, …]

"car" → [-0.2, 0.9, -0.1, …]

Here, cat and dog vectors are close, while car is far away.

🏗️ The Role of Vector Databases

When enterprises deal with millions of documents, they can’t compare every vector manually.
That’s where vector databases like FAISS, Pinecone, Weaviate, and Qdrant come into picture.

They:

• Store embeddings efficiently.
• Support Approximate Nearest Neighbor (ANN) search for speed.
• Provide APIs to integrate with LLMs for RAG pipelines.

🧑‍💻 Hands-On Demo with Python

Let’s see how embeddings + similarity search work in practice.

# Install dependencies:
# pip install openai faiss-cpu

import openai, faiss, numpy as np
openai.api_key = "YOUR_OPENAI_API_KEY"

docs = [
 "Cats are small domesticated animals that like milk.",
 "Dogs are loyal and bark at strangers.",
 "Cars are used for transportation on roads."
]

# Get embeddings from OpenAI
def embed(text, model="text-embedding-3-small"):
 resp = openai.embeddings.create(input=[text], model=model)
 return np.array(resp.data[0].embedding, dtype=np.float32)

doc_embeddings = [embed(d) for d in docs]
dim = len(doc_embeddings[0])

# Build FAISS index (cosine)
index = faiss.IndexFlatIP(dim)
normed_docs = np.array([v/np.linalg.norm(v) for v in doc_embeddings])
index.add(normed_docs)

# Query
query = "What animal likes milk?"
q_emb = embed(query) / np.linalg.norm(embed(query))

# Search
scores, idx = index.search(np.array([q_emb]), 2)

print("\n🔹 Results:")
for i, doc_id in enumerate(idx[0]):
 print(f"Rank {i+1}: {docs[doc_id]} (score={scores[0][i]:.4f})")

Expected Output

🔹 Results:
Rank 1: Cats are small domesticated animals that like milk. (score ~0.9)
Rank 2: Dogs are loyal and bark at strangers. (score ~0.6)

🏆 Key Takeaways

1. Embeddings transform text into machine-understandable vectors.
2. Cosine similarity is the go-to metric for semantic search in AI apps.
3. Vector databases like Pinecone and FAISS make retrieval fast at scale.
4. Together, they power RAG pipelines, chatbots, and agentic AI workflows in enterprise environments.

The future of enterprise AI isn’t just about large language models.
It’s about combining LLMs with retrieval, reasoning, and tool orchestration — and at the heart of it all lies vector similarity search.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI