Document chunk boundaries
Review a chunk without the document preview
A simple reading exercise reveals whether a retrieved passage carries enough context. Ask a reviewer to interpret it before showing the original page.
In this article
Give the reviewer the evidence the model receives
Select passages from several document structures, including ordinary prose, a nested list, a table and an appendix. Show the reviewer the exact text and metadata that the answer service receives. Do not add a helpful page screenshot at the start.
Ask them to explain what the passage means and what they cannot determine from it. For a number, ask for the unit and period. For a rule, ask who it applies to and whether an exception is visible. For a procedure, ask what must already be true before the steps begin.
The exercise is not a test of the reviewer's guessing ability. Uncertainty is useful evidence that a passage needs more context or should not support a standalone answer.
Reveal the source and compare the interpretation
Open the original document at the relevant version and location. Compare the reviewer's interpretation with the source. Identify the exact heading, note, label or neighbouring sentence that changes the meaning.
Record whether that information was lost during extraction, splitting or context assembly. This helps the engineering team choose a correction instead of applying a broad instruction to make chunks larger.
Include a passage that is already self-contained. A good review should recognise when no expansion is necessary. Otherwise, the process can drift towards attaching whole pages to every result regardless of the cost or ambiguity introduced.
Challenge the proposed boundary rule
If the team proposes repeating headings, ask how it handles a new section and an appendix. If it proposes joining table continuations, ask how it distinguishes a continuation from a separate table. If it adds neighbouring chunks, ask what happens when the neighbour contains a conflicting version of the rule.
Review provenance alongside meaning. Can the passage be traced to a source version and location? Will the citation still resolve after a new index build? Does the added context inherit the source's permissions?
These questions connect the reading exercise to the operating system. A chunk that reads well but cannot be traced or restricted is not ready to become trusted evidence in an application.
Finish with examples the team can rerun
Keep the reviewed source, extracted representation, emitted passage and expected interpretation together. Use synthetic or approved material so the fixture can be retained safely. Note any tool version needed to reproduce the extraction.
Write the finding in concrete terms. "The second-page row loses the table's unit heading" is more useful than "chunk quality is poor". Include an example of acceptable output without forcing every future implementation to produce identical wording.
After the change, ask the reviewer to repeat the blind interpretation on the corrected passage and one related case. This verifies that the rule preserves meaning beyond the exact example used to design it.
Primary sources
Microsoft Learn: document structure extractionMicrosoft Learn: chunking methodsReferences checked 11 September 2026.