A retrieval system does not find answers. It finds pieces of your documents, and the model only sees the pieces it is handed. Cut a document in the wrong place and the answer ends up in two pieces, with half of it never retrieved. Make the pieces huge and nothing gets cut, but you pay to send the whole page every time. Chunking is where a RAG system quietly wins or loses, before any model runs.
What runs. Three real chunkers cut one fixed 914-word handbook. Fixed starts a new chunk every N words. Sentence packs whole sentences until the next one would pass N. Structure splits at headings, then paragraphs, then sentences, and repeats the heading at the top of every chunk it makes. Retrieval is BM25 (k1 = 1.2, b = 0.75) over the current chunks. Each of the 8 questions has a gold answer stored as character offsets, and the verdict is computed from those offsets: HIT when one retrieved chunk holds the whole answer, PARTIAL when retrieved chunks hold part of it, MISS when none touch it.
What it skips. One word is one token here. No model writes an answer; the sim stops at the words a model would be given. Retrieval is keyword only, so an embedding search would change the scores but not the boundary problem.
The Harrow Ridge Stove and its handbook are fictional.