Episode 38: What Is Cosine Similarity? A Gentle Explanation of How AI Measures “Similarity” Between Texts

Measuring “Similarity” with Numbers? What Cosine Similarity Means

“This term sounds familiar, but I’m not quite sure what it means.” For many people, “cosine similarity” might be one of those phrases. The name itself may sound a bit mathematical, but the idea behind it quietly plays a role in our everyday lives.

For example, when a search engine suggests, “Is this the information you’re looking for?”—cosine similarity might be working behind the scenes to make that happen.

How AI Finds Similarities: Thinking in Arrows

So what exactly does cosine similarity do? Simply put, it’s a way to measure how similar two things are using numbers. But here, “similar” doesn’t refer to appearance or meaning directly—it’s about how close their directions are once they’ve been converted into numbers.

To make this easier to picture, imagine arrows. Data like text or images is turned into something called a “vector,” which is basically a collection of numbers. You can think of each vector as an arrow—it has both direction and length.

If two arrows (vectors) point in nearly the same direction, they’re considered very similar. If they point in very different directions, they’re seen as unrelated. What matters here is the angle between them—not their length. This means even short texts and long texts can be compared fairly using this method. That’s one of its key strengths.

Used in Search and Chatbots—But Not Without Challenges

This concept isn’t just limited to search engines—it’s also widely used in AI-generated writing and chatbots. For instance, when you see “recommended articles” or “related questions,” what’s happening behind the scenes is that each piece of text has been turned into a vector. Then their directions (i.e., content) are compared to judge how closely related they are.

Nowadays, combining this approach with large language models (LLMs) and vector search technology has become common. In these cases, techniques like “embedding” are used to convert text into vectors first. Then cosine similarity measures how close those vectors are—in other words, how similar the content is.

That said, there are some weaknesses too. For example, if two sentences mean almost the same thing but use completely different words, cosine similarity might wrongly judge them as unrelated. Also, when vectors have many dimensions (lots of numbers), it becomes harder for humans to intuitively understand what’s going on—so interpreting them requires care.

A Simple Idea That Supports Complex AI Systems

Even so, cosine similarity remains an essential part of many AI technologies and information retrieval systems today. The simple idea of “looking at direction” helps uncover closeness in meaning from massive amounts of data—a small but fascinating insight.

The information we receive every day often relies on quiet mathematical tools like this working behind the scenes. Cosine similarity is one such important mechanism playing its part.

In our next article, we’ll introduce GPUs—the unsung heroes powering AI computations. Why are they known for being good at image processing? We’ll explain that secret in an easy-to-understand way.