Checker

Chunkability Checker

Retrieval splits your page into fixed-size chunks before it ever sees a question. Where those splits land decides what gets quoted. This shows you the splits, and flags the chunks that arrive meaningless.

Ask about CreobotSee all tools

YOU GIVE ITa page and a chunk sizereadcomputereturnHOW THE PAGE SPLITSokokorphanYOU GET BACKwhere it splits and which chunks lose contextRuns in your browser. Nothing is sent anywhere.
  • Shows the actual splitsAt your chunk size
  • Flags orphan chunksThe ones that say nothing
  • Try different sizesSee where it breaks

Check it

Paste it here

Result

—chunks at this size
  • Orphan chunks —
  • Split mid-paragraph —
  • Smallest chunk —
  • Largest chunk —
  • Duplicated by overlap —
  • Verdict —

An orphan chunk is one that opens with a pronoun or a fragment, so a model reading it alone cannot tell what it is about. Those are the chunks that produce confidently wrong answers.

What the result means

Where the split lands decides the answer

A chunk that begins halfway through a paragraph reaches the model without its subject. It will still be quoted, because retrieval matched on the words that happened to be in it.

Overlap buys context and costs storage

Overlap repeats words across the boundary so a split sentence survives. Every overlapped word is stored and embedded twice, so it is a trade rather than a free win.

Short sections chunk better than long ones

If your paragraphs are shorter than the chunk size, splits land on paragraph boundaries and every chunk is a complete thought. That is worth more than any retrieval tuning.

Test at the size your vendor actually uses

Defaults vary between roughly 256 and 1,024 tokens. A page that chunks cleanly at one size can fall apart at another, so check the one you are on.

Questions

Whatever your vendor defaults to, unless you have a reason. The bigger win is writing sections shorter than the chunk so the splits fall on your boundaries.

No. It reduces orphans but stores and embeds the same words repeatedly, and it can make retrieval return several near-identical chunks instead of several useful ones.

This counts words because you can see them. A token is roughly three quarters of a word in English, so 512 tokens is about 380 words.

Want this run against your whole site?

Tell us the page your assistant lives on and the questions you care about. A person runs them and sends back the raw answers.