As discussed, traditional sequence-to-sequence models often compress the entire input sequence into a single fixed-size context vector. This approach can struggle to retain information from longer sequences, creating an information bottleneck. The attention mechanism provides a more flexible alternative, allowing the model to look back at the entire input sequence and selectively focus on the parts most relevant for generating the current output.

To formalize this selective focus, the attention mechanism employs an abstraction based on three components: Queries, Keys, and Values. Imagine you're researching a topic in a digital library.

Your Query ( $Q$ ) is the specific question or topic you're currently interested in.
Each document in the library has associated Keys ( $K$ ), which are like concise summaries or keyword lists representing the document's content relevant to potential queries.
Each document also has its actual content, the Value ( $V$ ), which is the detailed information you want to retrieve if the document is relevant.

The attention mechanism works similarly: for a given Query, it compares the Query against all available Keys to determine how well they match. This matching process generates a set of scores, often called attention weights. These weights indicate the relevance of each Key (and its corresponding Value) to the Query. Finally, the mechanism computes an output by aggregating the Values, weighting each Value according to its calculated attention score. Values corresponding to Keys that strongly match the Query contribute more to the final output.

In the context of neural networks, Queries, Keys, and Values are represented as vectors derived from the model's internal representations (embeddings or hidden states):

Query ( $Q$ ): A vector representing the current context or element that needs information from the input sequence. For instance, when translating a sentence, the Query might represent the decoder's current state as it decides which word to generate next. It asks: "Which parts of the input sentence are most relevant right now?"
Key ( $K$ ): A set of vectors, one for each element in the input sequence. Each Key vector is designed to be compared with the Query vector. It essentially represents what kind of information its corresponding input element offers.
Value ( $V$ ): A set of vectors, also one for each element in the input sequence, often paired with the Keys. The Value vector contains the actual content or representation of its corresponding input element that should be aggregated if that element is deemed relevant by the Query-Key comparison.

Flow of attention using the Query, Key, and Value abstraction. The Query is compared against all Keys to compute weights, which are then used to form a weighted sum of the corresponding Values.

The core idea is that the Query interacts with the Keys to understand where to focus, and the resulting attention scores determine how much of each Value contributes to the final output representation. This output is a context-aware vector that summarizes the input sequence elements relevant to the specific Query.

These Query, Key, and Value vectors typically have dimensions denoted as $d_q$ , $d_k$ , and $d_v$ respectively. The compatibility between a Query and a Key is often computed using their vector representations, commonly via dot products, which requires $d_q = d_k$ . The dimension $d_k$ plays a significant role in how attention scores are calculated and scaled, as we will see shortly. The dimension of the Value vectors, $d_v$ , determines the dimension of the output context vector before any final transformations.

It's important to note that the source of these Q, K, and V vectors defines the type of attention. In self-attention, which is fundamental to the Transformer architecture, Q, K, and V are all derived from the same sequence. This allows different positions within a single sequence to attend to each other. In cross-attention, commonly found in encoder-decoder architectures, the Query might originate from the decoder, while the Keys and Values come from the encoder's output, enabling the decoder to focus on relevant parts of the input sequence. For now, we focus on the general QKV abstraction itself.

This Query-Key-Value framework provides a powerful and flexible way to model dependencies regardless of their distance in the input sequence, directly addressing the limitations of fixed context vectors. The next step is to examine the specific mathematical operations used to implement this comparison and weighting process, starting with the Scaled Dot-Product Attention mechanism.