Cosine similarity in RAG: how it decides what your model "remembers"

If your Retrieval-Augmented Generation (RAG) system returns irrelevant passages, the problem often sits at the retrieval step. Cosine similarity is the most common metric that decides which embeddings