- Akamai measured one production document-ingest system. With page-based splitting, 43% of stored claims in one document did not end in a full stop, a sign of fragments.
- The loss happens before any model reads the text, so a search can return a grammatical sentence that is incomplete or misleading.
- Ask your team whether ingestion keeps headings, joins split sentences and records references, and whether missing context is marked as missing.
A sentence cut from page 40 of a report no longer knows it came from page 40. It does not know which heading it sat under. It does not know what its pronouns meant. Akamai's October post documents faults in one production ingest system that prepares company documents for AI search. The faults lose context quietly, before any model reads the text.
The risk is a true sentence with its context removed
Most talk about AI errors focuses on invented facts. Akamai describes a different failure. The sentence is real and well formed, but the pipeline that prepared it for search has removed what it depended on. Search then returns it as though it were complete.
This is the idea worth holding. A retrieval system can be accurate about words and wrong about meaning. Nothing downstream can recover what the ingestion step threw away. Bad answers can therefore start at the filing clerk, not at the model.
Akamai's post reports measurements from one production document-ingest system. The numbers below come largely from one long document. They show how the failure works. They are not a survey of the industry.
What Akamai measured
Akamai looked at a 686-page document, of which 590 pages produced any text. Nearly three in four of those pages (73%) opened partway through a sentence. Slightly more (78%) closed partway through one. In one case a page break cut the word "standards" in two, and two separate model calls were each asked to pull a fact from one half.
Akamai counted a stored claim as a fragment when it did not end in a full stop. That is a proxy, not a hand-checked count of broken sentences.
When the same document was split by sentence instead, the figure fell to 1%. The reprocessing took place four days after the splitting logic was corrected. Sentence-based splitting also produced fewer pieces, not more. Akamai's example of a stored fragment reads: "total income at the rate as specified in any Central Act for relevant tax year."
How the damage happens
Akamai lists four specific faults, all variations of one general case: a piece stored without the heading that gave it meaning. First, a page boundary is treated as a sentence boundary. Second, parsers find headings, because that is how they spot where one block ends, and then discard the heading text.
Third, on web content, headings are worse than lost. A heading has no full stop, so a splitter joins it to the paragraph below. Akamai shows a stored quote, "EdgeKV Key limits A namespace holds up to 128 keys." No such sentence exists on the source page. Two headings and a paragraph had been fused into one assertion. On a single page Akamai examined, the system saved 11 headings as if each were a fact the document had stated.
Fourth, sentences point elsewhere. A sentence beginning "Irrespective of anything contained in section 67" is accurate but depends on text the system did not keep.
Those claims held 1,665 references, of which 493 were distinct. Akamai separates three actions that teams often blur. Recording a reference is free and safe. Resolving it is a retrieval step that can fail. Swapping the reference for the section's content means a model is editing a statute, and nobody downstream can tell.
The fixes are cheap and need no model
Akamai's five remedies all run at parse time, with no model involved. The parser keeps a running list of open headings and attaches it to every piece. It joins a sentence split across pages before it splits text into sentences. It stops the join when the next page opens with a heading. It records headings as metadata, not body text. It stores citations in a separate field and resolves them at query time.
The page-joining change alone moved the fragment rate from 43% to 1%. Akamai notes that the heading list needs no extra reading of the document, because the parser already walks it.
The post also summarises research on richer approaches. One cited study found that roughly 40% of sentences either could not be rewritten to stand alone or needed no edit. Contextual Retrieval, as the post reports, cut the top-20 retrieval failure rate by 49%. These are heavier tools. The post's point is that cheap structural fixes come first.
Sometimes the document simply does not hold the context. Akamai's advice is to write down that it is missing. A blank entry is an honest record that the source was silent, and the system should not invent a default to fill it.
Questions to put to your team
Ask where your document pipeline splits text. If it splits by page, ask for the share of stored claims that do not end in a full stop. That is the simple fragment measure Akamai used.
Ask whether headings survive into each stored piece, and whether citations such as "section 67" are recorded as dependencies. Ask what the system returns when context is missing: a gap, or a guess.
For headings fused into body text, Akamai describes a separate check. Look for stored sentences with a capital letter mid-string and no preceding full stop. Ask whether anyone has run it.
The lesson is that AI answers inherit the habits of the plumbing beneath them. A confident answer built on forgotten context is still forgotten context.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Akamai.
