Recently edited does not mean currently valid
A draft and an approved policy can cover the same subject. Upload time cannot determine authority. Technical synchronization transfers updates; business ownership determines validity. Microsoft's deletion and change guidance and NIST's risk framework support these distinct concerns. The workflow here is an editorial proposal.
Create a minimal document identity
Store content owner, approval state, effective interval, organizational scope and successor document. Keep publication time, effective date and indexing time separate. A circular published this month may apply next month. Current answers should select the current rule; future-dated questions need explicit time context. Keeping history does not mean mixing every version into current answers.
For more on this topic, read Data architecture for enterprise AI agents in the Liyan knowledge base.
Measure freshness according to impact
Track source-to-index lag for each content class. Safety procedures and internal newsletters need different review priorities. Expose missed synchronization targets to the content owner. For sensitive guidance, a visible limitation or temporary refusal may be preferable to a confident obsolete answer. Define this behavior before failures occur.
Follow a change end to end
Trace a controlled update from source through retrieval, citations and role-specific visibility. Repeat with a cancellation. Report ownerless sources, missing validity dates and synchronization failures as actionable items. Document count alone says little about whether the underlying knowledge is current and authorized.
Represent when a rule applies
A policy published today may only become effective next month. Keep publication, effective start, effective end and ingestion timestamps distinct. Historical questions may require historical rules rather than automatic selection of the newest file. Preserve approved historical access separately from current guidance, and ask for the relevant date when it changes the answer.
Assign ownership and document relationships
Record an accountable organizational role, approver, audience, publication state and review deadline. Establish succession so staff changes do not create orphaned documents. Ask what a new document supersedes and which related tables or instructions also need review. Validate the resulting collection with representative questions, not merely an individual file's formatting.
Prioritize review by the consequences of error
Agree review intervals and overdue behavior using impact, change frequency and source dependence. An overdue flag signals uncertainty, not proof of falsity; warning or withholding depends on the workflow. Re-run reference questions after important changes and test obsolescence, deletion and access revocation. Report unowned sources, unknown validity, broken links, conflicting versions and resolution times with accountable owners. Corpus growth is not evidence of freshness: the outcome is access to the applicable authorized source at the relevant time.
Run a policy-change exercise across the whole knowledge chain
Create a policy that is published now but becomes effective later, while keeping the previous rule applicable to earlier requests. Ask present, future and historical questions. Check that the selected version follows the relevant date and audience rather than the most recent ingestion timestamp. If a user's date is ambiguous and changes the answer, request clarification rather than choosing silently.
Have the content owner record the successor relationship, affected guidance and review responsibility. A travel-expense update may also require a rate table and application guide to change. Test representative questions across those linked documents before marking the collection coherent. Assign an organizational role and substitute owner so review does not depend on one person's continued employment. Agree what overdue review means for each risk group: a warning may suffice for low-impact material, whereas a consequential answer may need renewed approval before citation.
After release, test a deleted source, a revoked role and a broken citation. Inspect whether the final answer and cached results reflect the change, not merely the document listing. Report unowned material, unknown applicability, conflicting versions and unresolved links with next actions and owners. Prioritize high-impact and frequently used sources when review work accumulates. The evidence of freshness is that an authorized user can reach the applicable, traceable source for a specific question and time; counting new uploads or successful ingestion runs cannot establish that result.
Frequently asked questions
Is the newest file always authoritative?
No. Approval, effective dates and scope also matter.
Does engineering alone own freshness?
No. Content owners determine validity; engineering operates synchronization.
Sources
- MicrosoftChange and deletion detectionOfficial documentation
- NISTAI Risk Management FrameworkOfficial framework
Further reading in the Liyan knowledge base
- Liyan Data ParsData architecture for enterprise AI agentsLiyan’s article explains how data ownership, freshness and access permissions shape the context available to an agent.
- Liyan Data ParsConnecting AI answers to organizational documentsContinue with Liyan’s guide to document preparation and evaluating retrieval separately from answer correctness.