How you cut documents matters
Split a rule in half and retrieval can't save you.
Documents are split into chunks before they're embedded. Cut in the wrong place and a rule lands in two halves, neither of which makes sense alone.
Chunked by section
#1 Returns0.777
Most items can be returned within 10 days of delivery, from the Orders page. Electronics (phones, headphones, laptops, chargers) are different: they can be returned only within 7 days of delivery, and only if unopened. Opened electronics can be replaced if faulty, not returned.
#2 Damaged or wrong items0.661
Report a damaged or wrong item within 48 hours of delivery, with a photo of the item and the box. After 48 hours we can't accept damage claims. We pick it up free and send a replacement, or refund you if it's out of stock.
#3 Cancelling an order0.642
You can cancel free of charge until the order is dispatched. After dispatch, refuse the delivery at the door instead, and the refund starts once it returns to our warehouse.
Chunked every 120 characters
#1 Chunk 10.741
…Most items can be returned within 10 days of delivery, from the Orders page. Electronics (phones, headphones, laptops, c…
#2 Chunk 20.725
…hargers) are different: they can be returned only within 7 days of delivery, and only if unopened. Opened electronics ca…
#3 Chunk 40.657
…g days. Cash on delivery orders are refunded to your bank account in 5 to 7 working days, after you add it under Payment…
Real gemini-embedding-2 embeddings, 2026-10-21
Cutting every 120 characters split “Electronics (phones, headphones, laptops, c” from “hargers) are different: they can be returned only within 7 days”. Split on meaning instead: sections, paragraphs, headings.
Which chunking is safest for a help centre?