Retrieval Is Mostly Chunking

Swapping embedding models buys you a little. Getting chunking right buys you a lot. The attention goes the wrong way round.

Placeholder article. Every retrieval system starts with an argument about which embedding model to use. Very few start with an argument about chunking.

Why it dominates

If the answer to a question is split across two chunks, no retriever will find it whole. If a chunk contains four unrelated topics, its embedding means nothing in particular.

Structure beats size

Chunk on the document's own boundaries — sections, clauses, table rows — rather than on a token count that respects nothing.

Replace this placeholder article with your own writing.

Read next

17 Sep 2026

Your Model Is Not the Product

The model is perhaps a tenth of what you are shipping. The other nine tenths decide whether anyone keeps using it.

6 min read · 11 views

10 Sep 2026

Evaluation Sets That Lie to You

A held-out set is only held out if it was separated before you started making decisions. Usually it was not.

8 min read · 6 views

0 comments

No comments yet. Be the first.

Leave a comment

Never published.

Plain text only. Comments from visitors are reviewed before they appear.