Lab 4: Retrieval by Hand & the Vector Store

Score a question against every sentence, keep the top few. Then let a store do it — and see what happens when a document changes.

🗄️

Retrieval is: embed the question, score it against everything stored, keep the top k

Five sentences you can score by hand. A million chunks you cannot — that is what a vector database is for. It stores the vector and the text, and answers one question: which stored vectors are nearest to this one?

Section 1: Retrieval, by hand

Edit the sentences or the question if you like. The question deliberately shares no words with the answer.

The stored sentences

The question

Keep top k =you choose k — the model has no say

Section 2: The vector store

Same sentences, now stored as records — id, vector, text. Query it. Then add a new version of a policy line and see what comes back.

Collection "policy" — 0 records

Empty. Press 1 to store.

operation log

Takeaway: Retrieval = embed the question, score it against everything stored, keep the top k — and you set k. A vector database does this at scale with an index, keeps the original text beside each vector (you cannot get it back from the numbers), and adds — never replaces. Next class: chunking a real document, choosing k, and handing the top chunks to the model.