OpenAI Embeddings Explained: Semantic Search, RAG & Vector Databases

Vector embeddings have revolutionized how software systems understand human language. In this 2026 guide, we cover how to get and optimize OpenAI embeddings (V3) for semantic search, retrieval-augmented generation (RAG), and vector databases.

To understand how to merge vector searches with keyword lookups, read our Complete RAG Hybrid Search Guide. Or, scale up your systems with our step-by-step LangGraph Multi-Agent Tutorial.

1. What are OpenAI Embeddings?

OpenAI’s text embeddings measure the relatedness of text strings. In simple terms, they translate natural language into list coordinates (vectors of floating-point numbers) representing semantic concepts.

The distance between two vectors measures their semantic similarity. When vectors are close together in vector space, they represent high similarity. When they are far apart, they represent low similarity.

2. How Embeddings Work

To get an embedding, send your text string to the embeddings API endpoint along with the model ID (e.g. text-embedding-3-small). The API outputs a normalized vector representing the text:

import OpenAI from "openai";
const openai = new OpenAI();

const embedding = await openai.embeddings.create({
  model: "text-embedding-3-small",
  input: "Your text string goes here",
  encoding_format: "float",
});

console.log(embedding.data[0].embedding);
// Output: [-0.006929, -0.005336, -0.000045, ...]

For third-generation embedding models, use the cl100k_base token encoding. Use a local library like tiktoken to calculate token lengths before making API calls to stay within rate limits and optimize costs.

3. Vector Database Integration

To perform search at scale across millions of vectors, you need a dedicated **vector database**. Popular choices include:

• Pinecone: Managed cloud-native vector database designed for high performance.
• ChromaDB: Lightweight, open-source embedded vector database for fast prototyping.
• Weaviate: Open-source vector search engine with hybrid search support.
• pgvector: PostgreSQL extension for indexing and searching vector data directly inside SQL.

4. Embeddings vs Fine-Tuning vs RAG

It is important to select the correct approach when building LLM applications. Here is how they compare:

5. Real Code Example: Calculating Similarity

Because OpenAI embeddings are normalized to length 1, calculating cosine similarity is simplified to a simple dot product. Let's compare "Apple iPhone" and "iPhone by Apple" in Python:

import numpy as np
from openai import OpenAI

client = OpenAI()

def get_embedding(text, model="text-embedding-3-small"):
    return client.embeddings.create(input=[text], model=model).data[0].embedding

# Generate embeddings
emb1 = np.array(get_embedding("Apple iPhone"))
emb2 = np.array(get_embedding("iPhone by Apple"))

# Calculate cosine similarity using dot product
similarity = np.dot(emb1, emb2)
print(f"Semantic Similarity: {similarity:.4f}")
# Output: Semantic Similarity: 0.9412

6. Reducing Embedding Dimensions

A key advantage of OpenAI's V3 models is the ability to truncate the vector length using the dimensions parameter without sacrificing performance. This reduces index storage requirements in your vector databases by up to 66%, boosting cost efficiency.

7. Frequently Asked Questions (FAQs)