𝗚𝗲𝗲𝗸 𝗢𝘂𝘁 𝗧𝗶𝗺𝗲: Embeddings -Continuing from “Chunk Happens” (RAG)

In our last Geek Out Time we poked at chunking — the unglamorous decision that quietly controls what even has a chance to be retrieved. This time we zoom in on the sibling that decides whether those chunks are actually found: embeddings. If chunking is how knowledge gets in , embeddings are how it’s remembered and located later.

I have run dense (semantic), sparse (keyword), and hybrid (both) retrieval in the Google Cola, same as before. Along the way I will include reranking and diversification — and show how small defaults like title+body weighting and a sensible α blend make a visible difference.

Background, minus the algebra

Think of an embedding as a call number for meaning. It’s a long vector that places a passage in a space where similar ideas sit near each other, so “sign-in error with Gmail” lives near “email login failing.” Dense retrieval uses those positions to find neighbors by meaning.

This is an excerpt — the full article continues on Medium.