An AI knowledge base does not become reliable simply because it contains many documents. It becomes useful when sources have clear identities, validity and ownership. A folder containing an old quotation, a new price list and informal notes without distinctions can produce persuasive but incorrect information. Cleanup should precede ingestion and continue as organisational knowledge changes.

Define which questions the collection must answer

Start with a specific audience and task. A collection for employee onboarding needs different material from one used for public customer service. Record common questions and the source an experienced person currently uses. If no approved answer exists, that is a content gap to resolve rather than an invitation for the model to improvise.

In a hypothetical example, an installation company wants an assistant that explains maintenance procedures by product. Its folder contains a first-edition manual, a newer revision and an internal email describing a temporary exception. The team must decide which source applies to which product and period. A recent email date does not automatically give its content general validity.

Use stable identifiers in the inventory

Give each source an identifier independent of its filename. Record its title, type, owner, version, scope, review date and permissions. Include its relationship to earlier or later versions. A file named “new final” does not establish whether it entirely replaces an old instruction or only one paragraph.

In the example, a procedure for a particular equipment model must remain distinct from general advice. Record whether a document concerns another market or installation. That information can later limit retrieval. If it is not captured during preparation, the system will attempt to infer the scope from wording. These small inventory decisions affect whether the resulting assistant can consistently select applicable information.

Separate copies from meaningfully different versions

Exact duplicates can be identified through technical comparison. Nearly identical versions require examination of changes. Two documents may differ in a single number that changes the entire decision. Do not simply retain the largest or newest file. Ask the responsible person to confirm the validity of differences.

Record a decision to retain, merge or retire each source. Removing something from the active collection does not always require destroying the historical file. It can remain in a separate archive with restricted use. Active knowledge and historical records need different rules so an old instruction is not presented as today's answer. Preserve enough context to explain why a replacement occurred.

Examine content as the system sees it

Test text extraction from PDFs, presentations and scanned documents. Check whether columns, footnotes or relationships between tables and titles disappear. A visually clear page can become confusing text after extraction. If a warning moves away from the instruction it limits, the answer may acquire dangerous certainty.

Splitting material into smaller passages must preserve necessary context. An amount requires its currency, period and conditions. A maintenance step requires its model and warning. Review samples of the passages that will actually be retrieved, rather than only the original file. AI Knowledge connects this preparation with retrieval design and evidence-backed answers. A successful extraction should retain the meaning needed to apply the instruction correctly.

Conflicts need a resolution rule

Define which source takes precedence by category and who decides when no clear rule exists. An approved policy may override an old presentation, but the business must establish that relationship. When a conflict remains unresolved, abstention and human referral may be the correct response. The model should not invent a compromise that no source approves.

In the hypothetical example, a temporary exception may apply only to a specific serial number. The team records that scope instead of ingesting it as a general rule. Include a question that triggers the conflict during testing. The answer should use the correct source or explain precisely why confirmation is needed. This tests the decision rule rather than merely checking that both documents can be found.

Maintenance should follow business changes

Assign an owner and review triggers: a new product, policy change, replacement instruction or expiry of a temporary exception. Retain a meaningful review date rather than simply the last file modification. When a source changes, check that the active collection updated and old content disappeared from current answers.

Acceptance criteria include correct sources for representative questions, separation of active and historical material, and abstention on unresolved conflicts. Include a question from a user without access permission. AI Consulting helps agree these rules before connecting knowledge to everyday operations. Maintenance responsibilities should be understood before launch so corrections do not depend on discovering errors through customers.

Frequently asked questions about knowledge cleanup

Can we keep every old version?

Yes, in an appropriate historical archive. It must be distinguished from active sources and should not be retrieved as a current instruction without a specific purpose. Historical availability and present applicability are different requirements.

Is the modification date sufficient?

No. It may reflect a formatting change alone. You need a review date, validity scope and a clear replacement relationship when a new version exists. These details help distinguish actual knowledge changes from file activity.

Who resolves conflicting instructions?

The knowledge owner, using the process established by the business. A technical specialist can identify the conflict but should not independently decide business policy. Record the decision so it can be applied consistently afterwards.

How do we know cleanup was sufficient?

Test real questions against verified answers and sources. If the collection supports the selected scope and handles gaps correctly, limited use can begin with ongoing review. Expansion should follow evidence from that initial scope.

For an organised knowledge base, contact DIGITALNOW. For further reading on information quality and reliability, see the NIST AI risk framework.

DIGITALNOW EDITORIAL TEAM

Practical guidance from DIGITALNOW, part of VNG Digital Group.

Editorial policy