# Understanding Embeddings : How AI Learns Meaning from Text, Images, and Data

![A Complete Guide to Creating and Storing Vector Embeddings! | by Pavan  Belagatti | Level Up Coding](https://miro.medium.com/v2/resize:fit:1400/1*QZFroonLyTxtOaA6AAImmw.png align="left")

**Hey Everyone** 👋🏻 **!**

In my previous blog, we explored how **Transformers work** and how they **revolutionized modern AI**, paving the way for major advancements.

Today, let's dive into **embeddings**—the **foundation of Large Language Models (LLMs)**. We'll cover **how embeddings work, different types of embeddings, and their applications.** Let's get started!

### **What is an Embedding?**

An **embedding** is a **numerical representation** of data (words, sentences, images, videos, audio, documents) in a **vector format** within a **multidimensional space**. These representations **capture meaning and relationships** between data points.

It is also known as a vector embedding.

📌 **Example:**

![Word Vector Mathematics Concept
(image by author)](https://miro.medium.com/1*BJ9yksA-xmubGhrIdAybDQ@2x.png align="left")

*Image sources :* [towardsdatascience.com](https://towardsdatascience.com/whats-behind-word2vec-95e3326a833a/)

  
Words with similar meanings—like **"king" and "queen"**—will have embeddings that are **closer together** in vector space.

---

## **Types of Embeddings**

1. ### **Word Embeddings**
    

* Represent individual words in a multi-dimensional space.
    
* Capture relationships between words.
    

📌 **Models:** Word2Vec, GloVe, FastText (by Facebook)  
📌 **Use Cases:** Machine translation, chatbots, search engines

2. ### **Text Embeddings**
    

* Represent longer texts (phrases, sentences, paragraphs, or documents) as vectors.
    

📌 **Models:** BERT, Paragraph2Vec  
📌 **Use Cases:** Text classification, sentiment analysis

> **Note:**
> 
> * **Word embeddings** focus on **individual words**.
>     
> * **Text embeddings** capture the **meaning of entire texts**.
>     

3. ### **Sentence Embeddings**
    

* Represent entire **sentences** in vector form.
    
* Capture **both meaning and context**.
    
* Similar sentences have embeddings that are **closer together**.
    

📌 **Models:** Sentence-BERT (SBERT), Universal Sentence Encoder (USE) \[by Google\], Infersent  
📌 **Use Cases:** Semantic search, text retrieval

> **Note:** Sentence embeddings are a type of **text embedding** but specifically focus on **entire sentences**

4. ### **Image Embeddings**
    

* Convert **images** into **feature vectors**.
    
* Helps in **image similarity searches & object recognition**.
    

📌 **Models:** CNN (ResNet, VGG, CLIP)  
📌 **Use Cases:** Image search, object detection

5. ### **Graph Embeddings**
    

* Represent **nodes, edges, or entire graphs** as vectors.
    
* Used in **social networks, fraud detection, recommendation systems**.
    

📌 **Models:** Node2Vec, GraphSAGE  
📌 **Use Cases:** Fraud detection, social network analysis

6. ### **Video Embeddings**
    

* Convert **both spatial (image) and temporal (motion) features** into a **meaningful vector sequence**.
    

📌 **Models:** C3D, CLIP (for video)  
📌 **Use Cases:** Video search, activity recognition

7. ### **Audio Embeddings**
    

* Convert **sound waves into vector representations**.
    
* Capture **pitch, tone, and speech meaning**.
    

📌 **Models:** **Wav2Vec, OpenL3**  
📌 **Use Cases:** Speech recognition, music classification

> **Note:** **All these embeddings fall under vector embeddings.**

---

## **Shared Embedding Space**

A **shared embedding space** is a common vector space where **different types of data** (e.g., text & images) are mapped **close together** if they are related.

![](https://www.dailydoseofds.com/content/images/2024/12/image.png align="left")

*Image sources :* [dailydoseofds.com](https://www.dailydoseofds.com/a-crash-course-on-building-rag-systems-part-5-with-implementation/)

📌 **Example:** **CLIP (Contrastive Language-Image Pretraining)**  
\- Developed by **OpenAI** to create a shared embedding space for **images & text**.  
\- Allows images and textual descriptions to be **compared directly**.

> **Use Case:** You can **search for images using text descriptions!**

---

## **Applications of Embeddings :**

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1738759812131/6ebb539c-290f-49eb-9bbf-646e567400f8.jpeg align="center")

*Image sources :* [lyzr.ai](https://www.lyzr.ai/glossaries/contextual-embeddings/)

* **Large Language Models (LLMs):** Convert input tokens into **token embeddings**.
    
* **Semantic Search:** Retrieves **similar sentences** to improve **search relevance**.
    
* **RAG (Retrieval-Augmented Generation):** Uses **sentence embeddings** to retrieve relevant text.
    
* **Recommendations:** Finds **similar products** using **vector search**.
    
* **Anomaly Detection:** Identifies unusual patterns in data.
    

---

## **Famous Word Embedding Models :**

![](https://miro.medium.com/v2/resize:fit:1400/1*SaTpzUFhBIFW71Dxy_49Vw.png align="left")

*Image sources :* [medium.com](https://medium.com/@punya8147_26846/unlocking-the-power-of-vector-embeddings-a-beginners-guide-to-their-types-and-applications-3b092a49516c)

* **Word2Vec** → Predicts a word based on surrounding words (**developed by Google**).
    
* **GloVe** → Similar to Word2Vec but with a different mathematical approach (**by Stanford**).
    

---

### Lets Understank Tokens :

![Understanding LLM Billing: From Characters to Tokens | Eden AI](https://cdn.prod.website-files.com/61e7d259b7746e3f63f0b6be/6630e466c569a5f73cd81c9e_Understanding%20LLM%20Billing_%20From%20Characters%20to%20Tokens.jpg align="left")

*Image sources :* [edenai.co](https://www.edenai.co/post/understanding-llm-billing-from-characters-to-tokens)

### **What is a Token?**

A **token** is a small unit of text used in NLP models. It can be:

* A **word** (e.g., "cat")
    
* A **subword** (e.g., "play" and "ing" in "playing")
    
* A **character** (e.g., "C", "a", "t")
    
* A **symbol or punctuation** (e.g., "!")
    

> **Example:**  
> \- Sentence: `"I love AI!"`  
> \- Tokens: `["I", "love", "AI", "!"]`

### **Token vs. Embedding**

| **Feature** | **Token** | **Embedding** |
| --- | --- | --- |
| **Definition** | A unit of text (word, subword, character) | A numeric vector representing meaning |
| **Format** | Text | Numbers (vector) |
| **Example** | `"cat"` → Token | `"cat"` → `[0.23, 0.87, -0.45, ...]` |

> * **Tokens** help models read text.
>     
> * **Embeddings** help models understand meaning.
>     

---

**Example Code: How Embeddings Works :** [https://github.com/Raghul-M/GenAI/blob/main/Embedings/Token-Embeddings.ipynb](https://github.com/Raghul-M/GenAI/blob/main/Embedings/Token-Embeddings.ipynb)

## Conclusion :

![The Matrix Recoded: Fan Fiction Movie Sequel Pitch | xcaliber | Commaful](https://media.tenor.com/images/e5c21d98f56c4af119b4e14b6a9df893/tenor.gif align="center")

Embeddings have become a fundamental building block in modern AI, enabling machines to understand and represent complex data in a way that drives advancements in natural language processing, search, and more. Their ability to convert words, sentences, or documents into numerical vectors allows for more efficient and accurate tasks like similarity search and classification. As technology evolves, embeddings will continue to play a crucial role in shaping the future of AI.

Connect with me on Linkedin: [**Raghul M**](https://www.linkedin.com/in/m-raghul/)