Business memory needs more than search.

Documents are not the same as knowledge

Two files may answer the same question differently. One may be outdated and the other may apply only to a particular customer. Knowledge architecture needs to account for those relationships. The goal is not to copy every file into one repository, but to make valid context available under control.

Provenance belongs with the answer

An answer supporting a business decision should lead back to the relevant sources. A filename may not be sufficient. Version, ownership and applicable context help with review. Where this information is missing, the claim must remain appropriately limited.

Easier search must not create a new access gap

A shared assistant must not automatically expose all information to everyone. Client separation and role permissions need to remain effective at retrieval. We therefore start with a bounded knowledge space, known sources and defined users.

Learning is a governed process

Feedback, storage and model training are not the same. By a learning company, we first mean documenting reviewed experience, improving rules and updating the knowledge foundation. Not every conversation becomes a new truth without review.

Example: which delivery terms apply to this customer?

In an illustrative knowledge project, the sources include a general price list, an older quotation and a current customer-specific agreement. A useful answer must distinguish the customer, validity period and authority of those sources. “I found three relevant documents” is not yet a resolved answer. If access to the governing agreement is missing, the assistant must not fill the gap from a similar customer’s file.

A small test set to start with

• Known question: does the system return the right answer and the relevant source passage?
• Conflict: does it identify different versions and what needs clarification?
• Missing evidence: can it explicitly say that the answer is unsupported?
• Unauthorised user: do restricted details stay hidden when questions are rephrased?
• Update: after the agreed processing step, does a superseded source stop being treated as current?

A knowledge graph can represent relationships such as “applies to”, “replaces” or “owned by” explicitly. Whether one is needed depends on the knowledge problem. For a bounded document collection, search, metadata and proper access checks may be a more appropriate starting point. We make that choice against the test cases, not a technology label.