RAG is mostly retrieval

The generation half of retrieval-augmented generation gets the attention because it’s the part that talks back. But the quality of the answer is decided upstream, in chunking, embedding, and ranking — long before a token is produced.

Relevance is a cosine similarity of normalised vectors, $\cos\theta = \frac{a \cdot b}{\lVert a\rVert\,\lVert b\rVert}$ — rank by it, take the top-$k$, and most “model” problems turn out to be index problems:

A tiny retriever, in practice:

import numpy as np

def top_k(query_vec, doc_vecs, k=5):
    q = query_vec / np.linalg.norm(query_vec)
    d = doc_vecs / np.linalg.norm(doc_vecs, axis=1, keepdims=True)
    return np.argsort(-(d @ q))[:k]

Placeholder body. The real piece walks through a small, honest index.