05 Chunks that are too big

5 / 11

Goal

Ask do gift cards expire and get a top result of 60 words or fewer that contains the word expire.

Chunk size is a trade-off. Big chunks carry plenty of context, but their vectors average many topics together — and whatever you retrieve, you pay for in the model's context window. Small chunks are precise, but may cut an answer away from what explains it.

Here the pages are cut into 120-word chunks, so asking about expiry retrieves the whole gift-card page: amounts, how to buy them, balances. The answer is in there somewhere.

Try it

Re-chunk smaller, re-embed, and ask again. Every chunk moves on the map, because every chunk is new.

Go to the terminal

Hint

The documents are cut into 120-word chunks, so the answer arrives buried in the whole page. chunk --size 30, embed, search again.

Show a solution

search "do gift cards expire"
chunk --size 30
embed
search "do gift cards expire"

This interactive environment needs JavaScript. Everything you need to read is on this page, and the solution works the same way for real.

RAG

Starting RAG…