Document chunk boundaries
Trace a misleading answer back to its chunk
Before changing the prompt, inspect the evidence the model received. A missing heading or broken table can turn accurate source text into a misleading answer.
In this article
Preserve the answer and its source versions
Capture the question, answer, request identifier and the passages selected for that request. Record the document version and ingestion configuration. The current source file may have changed since the answer was generated, so opening the latest document is not always a valid reproduction.
Keep the original extracted representation if it is retained by the ingestion process. That allows the investigation to separate a parser error from a splitting error. Without it, a team may need to rerun extraction using a tool version that is no longer installed.
Use the approved evidence location for sensitive documents. Debugging a misleading answer should not create an uncontrolled copy of the underlying customer material.
Compare three representations in order
First inspect the original page. Identify the heading, table labels, note or neighbouring paragraph needed to answer the question correctly. Next inspect the extracted document structure. Finally inspect the chunk and assembled context sent to the model.
This comparison reveals where meaning was lost. If a unit is absent from extraction, the parser or source quality needs attention. If extraction contains it but the chunk does not, the boundary rule is the likely cause. If the chunk contains it but context assembly discards it, the serving path needs investigation.
Also check whether passages from different source versions were combined. A correct paragraph from one revision and a correct table from another can produce an answer supported by neither complete document.
Contain the affected class of answers
If the issue is limited to a document type, consider temporarily excluding that type from generated answers while retaining a link to the source for manual reading. If a particular report is malformed, restrict that report rather than making an untested global chunk-size change.
Explain the limitation in terms the user can act on. For example, the application can say that the table needs to be checked in the source document. It should not present a confident calculated value when the required unit or scope is unavailable.
Preserve the failing fixture before reprocessing the collection. Reingestion may correct the visible problem, but the team still needs a way to demonstrate why the new output is safer.
Repair the narrowest responsible stage
Change extraction, splitting or context assembly according to the evidence. Compare the corrected passage with the old one. Verify that the required relationship is present and that unrelated material has not been added unnecessarily.
Run a small set of neighbouring structural cases, such as a table that spans pages and a genuine new table that starts on the next page. This catches a repair that solves one document by merging unrelated sections in another.
Record the parser and chunking versions used for the correction, then reprocess the affected source set through the normal publication path. The incident is not fully addressed if old malformed chunks remain available in an answer cache or a secondary index.
Primary sources
Microsoft Learn: document layout outputMicrosoft Learn: index rebuildingReferences checked 11 September 2026.