Where does the 120 to 220 word range come from?
The range is a working constraint adopted for the ftes.ai library, derived from what a self-contained passage has to contain rather than from a benchmark. A citable passage states a claim about a named entity, explains the mechanism behind it, and names a condition or limit. Written plainly in English that is three to six sentences, which lands between a hundred and twenty and two hundred and twenty words. It is honest to call this a constraint we chose and defend, not an optimum we measured against a retrieval system we do not control.
The value of publishing it as a range is that it converts a vague instruction — write self-contained sections — into something a writer and a script can both check.
What breaks below the floor?
Sections under roughly a hundred and twenty words almost always turn out to state a claim with no mechanism behind it. "The approval gate keeps a human in control" is a claim; without the next two sentences explaining that the work is held, presented with its evidence, and released only on a human decision, the passage asserts a benefit and proves nothing. Assembling models can use such a fragment only as a restatement, and restatements are replaceable by any other source that restates the same thing.
The ftes.ai content verifier enforces a fifty-word floor per section as a hard error for this reason. Fifty words is not the target; it is the line below which a section is definitely unfinished.
What breaks above the ceiling?
Sections over roughly two hundred and twenty words are usually two sections that were never separated. The tell is easy to spot: somewhere in the middle there is a sentence that could have been a heading. A passage covering two topics produces one vector that sits between them, which makes it a weaker match for either query than two focused passages would be. For a human reader the same problem shows up as a section that cannot be skimmed, because its first sentence no longer predicts its content.
Splitting is almost always the right move, and it is free: the second half already has its own proposition, so it only needs its own question-shaped heading.
How do you check your own sections?
Count the words per section, then read only the sections that fall outside the range. The count is mechanical and takes seconds; the reading is where the work is. For a section under the floor, ask which of the three parts is missing — the claim, the mechanism, or the limit — and write it. For a section over the ceiling, find the sentence that wanted to be a heading and split there. The ftes.ai library runs this check in its content verifier so that a piece cannot ship with an unfinished section.
A useful second pass is to read each section's first sentence in isolation, in order. If that list reads as a coherent summary of the piece, the sections are carrying their own propositions.
Does the range apply to every kind of content?
It applies wherever a passage might be retrieved alone, which is nearly everywhere: explainers, playbooks, comparisons, frameworks, and whitepaper sections all obey it in the ftes.ai library. The exceptions are structural rather than editorial. A glossary entry may be shorter because the whole entry is one definition and the entry is the retrieval unit. A checklist item is not a passage at all; the section that introduces the checklist is. A frequently asked question and its answer form their own small unit, which is why those answers are held to a different rule — at least two of them short enough to be spoken aloud.
The principle underneath is constant: whatever the reader or the retriever receives alone has to make sense alone.
Why not simply write longer pages?
Because length is not the quantity that competes. A four-thousand-word page made of six sprawling sections offers six retrievable units, several of them ambiguous. A fourteen-hundred-word page made of eight tight sections offers eight clean units, each a candidate for a different sub-question. The shorter page is the stronger source on a per-question basis, and it is the one a human finishes. Word count as a target produces padding, and padding is the one thing guaranteed to lower the density of every passage it is spread across.
The ftes.ai library sets a floor on length rather than a target for that reason: nine hundred words for a standard piece, two thousand two hundred for a whitepaper, and nothing to gain from exceeding either without more to say.
What this range does not claim
The range is not a measured retrieval optimum, and publishing it as one would breach the honesty rule this library is built on. No public retrieval system discloses its chunking strategy, its overlap, or its scoring, and those choices differ across systems and change without notice. What the range does is encode a property that holds regardless of implementation: a fragment that contains one complete idea is more useful than one that contains half an idea or two competing ones.
If a system's own chunking cuts across our sections, well-formed sections still degrade gracefully, because each sentence already names its subject. That graceful degradation is the real reason to hold the line.