Attention as Query–Key–Value Weighting and an Encoder–Decoder Bridge

Seq2seq with a single context vector forces every input token to compete for space in one fixed summary — a bottleneck when a long filing or price path contains several facts relevant at different output steps. Attention replaces that summary with a data-dependent weighted read from all encoder states. This note develops the query–key–value (QKV) formalism and the encoder–decoder attention bridge; it does not reproduce the full Transformer survey in Attention, Transformers, and Self-Supervised Foundation Models.