How do we tell whether this is our problem?

Look at what the system stored rather than at what it answered. Take a document with a table in it, find the passage the system holds for that part, and read it. If the numbers have arrived without their headings, you have found the stage, and no amount of work on search or on the model will improve those answers.

Which documents suffer most?

Anything whose meaning lives in its layout. Price lists, specifications, comparison tables, forms and presentations all carry their sense in the arrangement rather than in the sentences, so they degrade sharply. Continuous prose survives almost anything, which is why a corpus of policies feels fine while a corpus of datasheets does not.

Is it worth handling some file types differently?

Frequently yes, and it is the cheapest available improvement in many deployments. Treating spreadsheets as spreadsheets rather than as documents, and giving scanned material its own path, tends to recover more than a great deal of tuning elsewhere. The gain concentrates on exactly the documents people complain about.

Glossary

Document parsing

Document parsing reads a file's actual structure rather than only its words, and because everything downstream inherits what it produced, a table flattened here cannot be recovered anywhere later.

In plain terms

Working out what a file actually is: which lines are headings, which block is a table, which column belongs to which row. Get it right and the meaning survives. Get it wrong and a neat table becomes a run of numbers with nothing to attach them to, and no later stage has any way of knowing that those numbers ever had headings.

Why it matters

Because it happens first and everything else inherits the result. Cutting into passages, searching, ranking and answering all operate on whatever this stage produced, so a document that arrived as an undifferentiated wall of text is permanently degraded no matter how good the rest of the arrangement is. Nothing later can restore structure that was discarded here.

How it works

Structure has to be reconstructed rather than read off, because most formats record appearance rather than meaning. A heading is frequently just larger text, a table is a set of positioned cells, and a two-column layout is two streams that a naive reader will interleave line by line into nonsense, so the parser is inferring intent from geometry throughout.

Tables are where the damage concentrates, and they are common in exactly the documents that matter. A price list, a specification or a comparison flattened into a sequence loses the association between a value and what it describes, which produces a passage containing all the right numbers attached to nothing. An answer built from it will be specific, fluent and wrong.

File types vary enormously in how well they survive this. A document exported cleanly from a word processor keeps most of its structure, while a scanned page has none to begin with, a presentation is a set of boxes with no reading order, and a spreadsheet saved as a document loses the thing that made it a spreadsheet. Corpora are mixtures of all four.

The failure is silent, which is what makes it durable. There is no error and no flag; the material simply arrives less structured than it was, and the consequence appears much later as an answer that is subtly wrong about a figure. Nobody traces that back to a stage they were never told existed.

The same table, before and after

The reason this stage escapes scrutiny so completely is that it produces no artefact anybody looks at. A team evaluating an internal search tool will read its answers, and if the answers are poor they will look at ranking, at the model, at how the question was phrased. What almost nobody does is open the stored passage for a document they know contains a table and read it as text, which takes two minutes and settles the question immediately. If the numbers have arrived stripped of the headings that gave them meaning, then every answer touching that document was built from noise, and the system was behaving impeccably throughout: it retrieved the right passage, handed it over faithfully, and produced the most plausible reading available of material that had already lost its sense. The defect happened before any of the parts people think about were involved.

Seen in the wild

Document intelligence placed inside the reader itself, which documents outright that images, complex vector graphics and scanned pages are outside what it handles.
Adobe Acrobat AI Assistant
A research tool answering strictly from uploaded material, where a reference that jumps to a passage depends on that passage having survived being read.
NotebookLM
A private document tool that handles taking files in and preparing them locally, so the quality of this stage is set by defaults on your own hardware.
AnythingLLM

Common misconceptions

People assume

It is the same as reading the characters off a page.

In fact

Character recognition produces text and this produces structure, which is a harder and more consequential job. A page can be transcribed perfectly and still arrive with its table flattened and its headings indistinguishable from body text, at which point the words are all present and their meaning is not.

People assume

A parsing problem will be obvious.

In fact

Nothing reports it. The material arrives looking like ordinary text, the pipeline proceeds normally, and the consequence surfaces later as an answer that is confidently wrong about a number. Because the stage is invisible to everybody using the system, the investigation almost never reaches it.

Telling them apart