How Recommendation Systems Work (No Jargon)

The Question

Have you ever wondered how websites like Amazon know exactly what products to recommend when you shop? Or how Netflix seems to suggest movies you end up enjoying?

The core question these companies solve every day is: How do we show each person things they'll probably like? Whether it's products on Amazon, videos on YouTube, or movies on Netflix.

This guide will walk through exactly how modern recommendation systems work, starting from the simplest approach and building step by step to the powerful vector-based systems used today.

Walkthrough: Our Tiny Store

Imagine we launch a 6‑item store. We try to maintain related lists by hand. Then we add a new item and realize how fragile this gets.

items = [banana, milk, bread, eggs, protein, shaker]
related = {
  banana: [milk, bread, eggs],
  protein: [shaker],
}
// Add "almond butter" → nothing recommends it; it recommends nothing
// You'd have to edit many lists by hand.

Attempt 2: Let Data Draw the Map

Let's improve our approach. Instead of manually creating lists, we can track what customers actually buy together and let the data guide us.

Every time two products are purchased together, we strengthen the connection between them. The more often they're bought together, the stronger their link becomes.

This approach works well and learns automatically from customer behavior. But it has problems at scale:

// Adjacency matrix example (co-purchase counts)
// rows/cols: [banana, milk, bread, beer]
[
  [0, 42, 31,  1],  // banana was bought with milk 42 times
  [42, 0,  15,  0],  // milk was bought with bread 15 times
  [31, 15, 0,   0],  // etc.
  [1,  0,  0,   0]
]
// New item "chia" → new row & column full of 0s (cold start)
// With millions of products, this matrix is gigantic

Attempt 3: Organizing Items on a Number Line

Let's try a different approach. What if we assign each product a single number based on some important feature? Then we could place all products on a simple number line.

The Multi-Dimensional Solution: Movies Have Multiple Traits

We've tried organizing items with one number, but it's not enough. Let's think about Netflix movies as a real-world example:

What if instead of one number, we give each movie several numbers - one for each important trait?

Mapping Items in Multi-Dimensional Space

Let's visualize how this works with a simple example. Imagine we track just two traits for simplicity: Action (0-1) and Funny (0-1). Each movie becomes a point on a 2D map.

How Netflix Actually Uses This Approach

Introducing Vectors and Embeddings

// Simple vector comparison
const A = [2, 2]; // direction is 45°
const B = [4, 4]; // same direction, 2x magnitude
const C = [4, 1]; // different direction

// A and B point the same way (similar movies)
// A and C point differently (different movies)

Vector Databases: Storing and Searching Efficiently

Try It Yourself: Vector Thinking

Now that you understand the concepts, try applying vector thinking to a simple example:

Summary: From Arrays to Vector Databases

We've covered a lot! Let's recap the journey we've taken from simple recommendation systems to modern vector databases: