# A table cell is not a complete answer

Splitting a document into searchable passages can remove the information that makes a sentence or number meaningful. Choose boundaries around the questions the source needs to answer.

By Cobnex editorial. Published 2026-09-10. Updated 2026-09-11.

## Read the passage without the page

A report contains a table headed "Monthly volume, thousands of units". One row contains the value 12. A search result retrieves the row label and the number but loses the heading. An assistant can now give a neatly cited answer that is wrong by a factor of a thousand.

The text was extracted correctly. The search found the relevant row. The failure is that the retrieved unit of information was incomplete. Similar problems occur when a paragraph loses a preceding exception, a recommendation loses its section heading or a figure loses the caption explaining what it measures.

A chunk is simply a piece of a larger document made available for retrieval. Its size matters, but its meaning matters more. A useful design question is whether someone could interpret the chunk correctly without opening the surrounding page.

## Preserve structure before enforcing a size limit

Begin with the document's logical structure where it is available. Headings, paragraphs, lists, tables and captions are more useful boundaries than an arbitrary character count. Layout extraction can help recover this structure from PDFs, but its output still needs checks against the original document.

Keep the heading path with each passage. A paragraph under "Existing customers / Exceptions / Regional delivery" means something different from the same words under a general policy heading. Preserve the source version and location too, so the answer can point back to the evidence actually used.

Large sections may still need splitting to fit model limits and keep retrieval focused. Apply the size limit after recognising the structure. Carry the necessary heading or table context into the smaller pieces rather than assuming the model can reconstruct it.

### From document to usable evidence

The pipeline keeps source structure and provenance before creating the passages used by search.

1. **Original document**: Retain the version and a viewable source
2. **Structural extraction**: Identify headings, tables, captions and reading order
3. **Bounded passages**: Split large sections while preserving their meaning
4. **Context selection**: Retrieve evidence with the labels needed to interpret it

## Use different treatment for different material

A short policy paragraph may work as a self-contained passage. A table row may need the table title, column headings and a note about units. A long procedure may need several steps together because a later instruction depends on an earlier condition.

This does not mean every document needs a bespoke parser. A small number of documented content rules can cover much of a collection. The important distinction is between deliberately supporting a structure and flattening everything into text while hoping overlap will repair it.

Overlap repeats some text between neighbouring chunks. It can help at a boundary, but it also increases duplication and prompt size. It is not a reliable substitute for a table header that was never extracted or a footnote whose relationship to the row was lost.

## Evaluate interpretation, not just retrieval

Use questions that depend on units, dates, exceptions and heading scope. Compare the selected passage with the source page and the final answer. A relevance score may look good even when the answer misinterprets a correctly retrieved number.

Include difficult documents in the fixture: a table that spans pages, a two-column report, a scanned appendix and a policy with nested exceptions. Keep the expected relationships explicit so a parser update can be judged against them.

There is no universal chunk size that fixes all of these cases. Start with a simple structure-aware approach, inspect its failures and add rules where the collection justifies them. The result should be evidence that remains understandable after retrieval, not merely text that fits inside a token limit.

## Sources

- [Microsoft Learn: document chunking approaches](https://learn.microsoft.com/en-us/azure/search/vector-search-how-to-chunk-documents)
- [Microsoft Learn: document layout analysis](https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/prebuilt/layout?view=doc-intel-4.0.0)