A document owner corrects a procedure, but the AI assistant continues answering from yesterday's indexed copy. Connecting a source once is not the same as keeping knowledge current.
Document synchronisation needs to handle edits, deletions, access changes, and failed processing as visible states.
Track the source and indexed version separately
For each source item, retain a stable reference and the version or change marker your connector can verify. Track the version successfully processed into the index as a separate value.
Do not mark an item current simply because the connector fetched it. Extraction, chunking, indexing, and permission updates can fail after retrieval.
Make replacement a controlled operation
For a fictional company handbook, an update might proceed like this:
| Stage | Evidence to retain |
|---|---|
| Change detected | Source identity and version |
| Content fetched | Retrieval outcome |
| Extraction completed | Usable text and error checks |
| New index content ready | Indexed generation reference |
| Active version changed | Current searchable generation |
Avoid deleting the old index entries before you know the replacement can be used, unless policy requires immediate removal. Conversely, do not keep both versions eligible indefinitely and hope ranking chooses the correct one.
The policy-version guide explains why explicit applicability rules still matter after synchronisation works.
Treat deletions and revoked access as first-class changes
A connector that only imports new files will eventually retain material that should no longer be available. Define how it detects removals and permission changes, including when notifications are missed.
Use periodic reconciliation where appropriate. Event-driven updates can reduce delay, but a missed event should not become permanent divergence.
If a source becomes unavailable, distinguish “not currently reachable” from “confirmed deleted.” The correct response may differ. For access-sensitive material, use the agreed fail-closed policy rather than silently serving stale restricted content.
Show stale or failed knowledge to its owner
Maintain a review view with the source, last successful sync, current error, and affected collection. Give document owners a useful way to fix unsupported files or missing permissions.
The user-facing assistant may need to explain freshness limits when the answer depends on a recently changed policy. Do not imply that an indexed answer is live source verification if it is not.
Test changes end to end
Use fictional documents. Edit a decisive sentence, delete a file, revoke access, rename it, and interrupt extraction. Ask the same question after each transition and inspect the retrieved source.
A full rebuild is simpler to reason about for a tiny collection but costs more as content grows. Incremental updates reduce repeated work while adding version and recovery complexity. Choose based on collection size and freshness needs.
My Nucleus case study provides context for this integration work, without claiming a specific synchronisation benchmark.
Knowledge-sync questions
Does fetching a new file mean the answer is current?
No. Verify successful processing and active retrieval state.
Can we rely entirely on notifications?
Consider a reconciliation path for missed or failed updates.
What should a pilot include?
An edit, deletion, permission revocation, and failed import. Use those cases in a knowledge integration review.