The Librarian’s Logic: Queries, Keys, and Values in Transformers

The defining breakthrough of the Transformer—the architecture powering every major AI today—is "Scaled Dot-Product Attention." But what does that actually mean?

To understand how Transformers process language, forget about reading word-by-word. Imagine a library where every word is a book, and the model is a librarian trying to find the context.

The Trio: Q, K, and V

For every word (or "token") in a sentence, the Transformer generates three unique vectors:

1. The Query (Q): What is this word looking for? (e.g., "I am a verb; where is my subject?")
2. The Key (K): What does this word have to offer? (e.g., "I am a noun; I can be a subject.")
3. The Value (V): What is the actual content of this word?

The "Attention" Score

When the model processes a sentence, it performs a "search" for every word simultaneously:

• It takes the Query of the current word and compares it against the Keys of every other word in the sentence (using a dot product).
• The more a Query and a Key "match," the higher the score.
• These scores are then "normalized" (via Softmax) to create weights that add up to 100%.

The Result: Contextual Meaning

The model then multiplies these weights by the Values of the corresponding words.

If the sentence is "The animal didn't cross the street because it was too tired," the Query for "it" will find a strong match with the Key for "animal." The final representation of "it" will then be heavily weighted by the Value of "animal."

This is how Transformers solve the "Ambiguity Problem." They don't just see words; they see how every word attends to every other word. It’s not just a leap in efficiency; it’s a leap in understanding context at a global scale.