Skip to main content

Q, K, V: The Library Search

We know that in Self-Attention, every word looks at every other word to figure out how they relate. But if you have 1,000 words all yelling at each other at a cocktail party, how does the AI actually organize the math?

It uses a brilliant system borrowed from database search engines: Queries (Q), Keys (K), and Values (V).

The Library Analogy​

Imagine you are in a massive library, trying to find a specific piece of information.

  1. The Query (Q): This is what you type into the search bar. It's what you are looking for. (e.g., "Books about space").
  2. The Key (K): This is the title and tags on the spine of every book on the shelf. It's what the book is. (e.g., "Title: Apollo 11, Tag: Space").
  3. The Value (V): This is the actual pages inside the book. It's the content you want to read.

How Words Use Q, K, V​

In a Transformer, every single word splits its personality into three different vectors: a Query, a Key, and a Value.

Let's look at the sentence: "The fluffy cat purred."

Take the word "purred". As a verb, it needs to find out who did the purring.

  • Query (Q) of "purred": "I am an action looking for a furry animal."
  • Key (K) of "cat": "I am a furry animal."
  • Value (V) of "cat": [The actual math meaning of the word cat].

The Matchmaker Math: The AI takes the Query of "purred" and multiplies it (using a dot product) against the Keys of every other word in the sentence.

When the Query "looking for animal" hits the Key "I am an animal", the math explodes into a huge match score (like 95%)!

Once "purred" finds its match ("cat"), it grabs the Value of "cat" and mixes it into its own meaning. Now, the AI knows that the purring was specifically done by the fluffy cat!

The "Scaled" Part of the Math​

If you read the original paper, the math is called Scaled Dot-Product Attention. What does "Scaled" mean?

Remember the "Exploding Gradient" problem from Chapter 3? If you multiply giant vectors together, the numbers can get so big that they break the AI's brain.

To fix this, after the AI multiplies the Query and the Key, it divides the answer by a scaling factor (usually the square root of the vector size). Think of it like turning down the volume on a speaker before it blows out your eardrums. It keeps the math stable and healthy!

Next Up: A single library search is great. But what if we want to search for grammar, emotion, and logic all at the exact same time? We need a team of detectives: Multi-Head Attention!