An AI assistant can confidently answer with a condition from last year's offer, even though the website already has a new version. A reliable knowledge base requires more than uploading documents: you need to know which source takes priority, when it's updated and whether the old information has actually been removed from retrieval.
Manage the knowledge base with one owner per topic, a canonical source, a version, a validity date and an update rule. After a change, check the index and the answers with test questions. For dynamic data, use a suitable system or API. When a source is missing or contradictory, the assistant should state the limitation and ask for a human check.
Split knowledge by how fast it changes
A process description can stay valid for a long time. A price, an available slot or an order status can change immediately. If everything is stored as static text, you risk the assistant turning yesterday's state into today's promise.
| Type of information | Suitable source | Control |
|---|---|---|
| Service descriptions | An approved page or document | Periodic review |
| Terms and limitations | A managed current version | Effective date |
| Prices and availability | The operational system, when an integration is available | Time of the check and error handling |
| Status of a specific customer | A protected system with suitable access | Verified identity and permissions |
| Historical materials | An archive separate from active knowledge | Not to be used for a current promise |
Create a source register
The minimum register contains an identifier, topic, URL or document, owner, version, validity date, last review and status. Add language and audience when conditions differ between markets or customer groups.
Sample entry: "Delivery policy; owner: operations; version 3; valid from [date]; public; active". This is a template, not an actual DigitaL+ policy. The goal is that the person fixing an answer can find the exact source and establish why it was used.
How to resolve contradictions
Don't leave two documents with different terms as equal sources. Define which one takes priority and why. A newer upload date doesn't prove a newer business policy: an old file may have been re-uploaded.
When active sources contradict each other, temporarily restrict the specific answer and send a task to the owner. The assistant can say the information needs clarification instead of choosing the more convenient version. For contractual or customer terms, applicability must be established from the real context, not by assumption.
Updating has several separate steps
- Approving the change in the primary source.
- Extracting the current content and metadata.
- Updating the documents and the corresponding parts in the index.
- Removing or excluding the replaced versions.
- Checking that retrieval returns the new information.
- Testing the final answer and the link to the source.
In RAG systems, information is retrieved from an index or other sources and used to generate an answer. Changing the original file doesn't guarantee the derived information has been updated. Microsoft describes separate operations for adding, updating and removing in an index; the specific process depends on the platform used.
If the document is split into chunks, replace all related chunks by a stable identifier. Adding a new version without removing the old one can leave contradictory excerpts available. Also check the caches that affect the answer.
Dynamic data is not an ordinary FAQ
For availability, use a check in the real system if the assistant has a permitted integration. If the request to it fails, don't return the last known value as a current fact. State the limitation and offer a suitable next step.
Don't let a public assistant retrieve internal documents just because they're in a shared index. Access must be checked before the information reaches the generated answer. A customer status must not be shown to anyone who knows an order number.
A test matrix for outdated answers
| Test question | Expected behavior |
|---|---|
| A question about a changed condition | A current answer and the correct source |
| A question with an old condition in the text | A correction, without accepting the premise |
| Two sources contradict each other | A limitation and a task for review |
| Not enough information | A clarification or a handoff to a human |
| A request for inaccessible internal data | Respecting the permissions |
| The availability API doesn't respond | No invented current availability |
Also use natural rephrasings, spelling mistakes and questions in the supported languages. Check the content, not just whether the answer sounds good. Citing a document is only useful if it actually supports the claim.
Who is responsible after launch
Assign a knowledge owner and a technical owner for updates. Track failed imports, the age of sources, test errors and user reports. Each problem should be linked to a source version and configuration whenever possible.
For B2B companies DigitaL+ combines AI assistants and chatbots with managed automations. When the information isn't sufficient, the process for handing a conversation over to a human is useful too.
Frequently asked questions
Does RAG eliminate inaccurate answers?
No. It helps use specific sources, but retrieval quality, freshness and the generation itself remain important.
How often should the knowledge base be updated?
According to the risk and the speed of change. For some documents a periodic review is enough; operational data requires a live check or a clear freshness boundary.
Is it enough to delete the old PDF?
Not necessarily. Check the derived chunks in the index, archive sources and caches the assistant may use.
