Document chunk boundaries
Rechunk a corpus with stable source references
Changing passage boundaries should not make citations point to unrelated text. Separate source identity from the temporary shape of an index build.
In this article
Distinguish a document from one of its passages
A document identifier should describe the source, while a passage identifier describes a particular derived piece of that source. When splitting rules change, the document can remain the same even though every passage identifier changes.
Include the source version and chunking version in the derived record. Avoid using a passage's ordinal position as though it were a permanent source reference. Passage 12 in a new build may contain entirely different text from passage 12 in the previous build.
For citations, retain a source location that can be resolved against the correct version, such as a page and section reference where the format supports it. If the application cannot display historical versions, make that limitation explicit rather than silently opening the latest file under an old citation.
Build a parallel derived collection
Run the new splitter into a separate index or versioned collection. Preserve the old serving path while inspecting the new records. This allows comparison without exposing a mixture of old and new chunks to ordinary queries.
Reconcile source coverage by document and version. A changed chunk count is expected, so count alone is not a success criterion. Check that each source produced usable passages, that failures are visible and that removed documents have not reappeared.
Carry access metadata through the rebuild. A chunking migration should not temporarily publish records without their source restrictions. The fact that the content was already searchable in another index does not make an incomplete new record safe to serve.
Compare evidence selection and citation resolution
Run questions that depend on headings, tables and cross-paragraph context against both collections. Inspect which source spans are selected and whether the answer remains supported. A new boundary can improve recall while making citations less precise.
Open the citations produced by both paths. Verify that the displayed source matches the document version and passage meaning used in the answer. Include a document that changed during the rebuild so version handling is exercised rather than assumed.
If conversations retain passage identifiers, decide how old references are resolved after cutover. A lookup table, retained old collection or clear source-unavailable state may be needed. Reinterpreting an old identifier against the new chunk sequence is unsafe.
Retire the old representation deliberately
At cutover, record the serving collection, chunking configuration and reconciled source checkpoint. Keep a rollback path only if it still respects current source content and permissions. An old index that stopped receiving deletions is not a safe fallback simply because it remains online.
Invalidate answer caches whose provenance depends on the old passage representation, or retain a verified way to resolve that provenance. Do not assume a cache expiry will occur before someone opens an old conversation.
After the agreed observation period, remove unused derived records and migration credentials. Keep the evaluation fixture and the mapping decisions. They explain how citations and stored conversations were handled and prevent the next rebuild from rediscovering the same problem.
Primary sources
Microsoft Learn: rebuilding an indexMicrosoft Learn: layout-based chunkingReferences checked 11 September 2026.