Knowledge quality and access boundaries

Home Learn Knowledge quality and access boundaries

In short

An assistant needs information that is relevant, current, and suitable for the question. It also needs to respect the requesting person’s right to access that information. Better wording can’t repair a source collection with unclear versions or permissions.

Why it matters

A team can connect an assistant to a document collection and inherit every problem in it. Duplicate policies, missing owners, and old instructions can all become answer material. Restricted documents can also become an exposure risk if the connection ignores access boundaries. Organizing the knowledge and controlling access are part of building a useful assistant.

How it works

The knowledge system needs an owner

Data quality means that information is fit for its intended use. Accuracy matters, but so do completeness, relevance, consistency, and currency. A correct instruction from last year can be wrong for today’s question. Provenance describes where information came from and how it was prepared.

Metadata is information about a source, such as its owner, status, effective date, and intended audience. Give approved material a clear identity and a retirement process. Preserve relationships between a policy and its exceptions. The Datasheets for Datasets research proposes documenting how data was collected, what it contains, and its intended uses. That same habit helps people understand the material feeding an assistant.

Access must follow the person

Authentication establishes who a user or service is. Authorization determines what that identity may do. NIST’s guide to attribute-based access control describes decisions based on characteristics of the subject, resource, requested operation, and relevant conditions. It makes the access decision a system responsibility.

The connected service needs to enforce those rules for the requesting identity. An assistant’s own service account mustn’t silently widen the information available to everybody. Check retrieved passages, snippets, citations, cached answers, and logs. A cache is stored information reused to avoid repeating work; it can preserve material after access changes if the system doesn’t handle that change.

Build a small trusted collection

Start with a narrow set of approved documents. Assign owners, remove superseded material, and create questions with known source-supported answers. Include requests that should be refused because the information is unavailable or restricted.

Separate a missing-information problem from a permission problem. One may need a better source; the other needs a denied operation. I’d review both when a source is removed, a role changes, or an account loses access. A prompt telling the assistant to be careful doesn’t replace enforced authorization.

Example

Illustrative example: this scenario is invented.

Daniel teaches adult-learning classes. His collection includes a public course outline and a restricted administrative note. He wants a learner-facing assistant to explain the outline.

The collection needs to supply the public material without exposing the administrative note. If the outline has been replaced, the old version should stop appearing as current guidance. Daniel checks the source identity and the access path before worrying about the answer’s tone. A readable explanation based on the wrong document still needs correction.

What it doesn’t mean

  • More documents don’t automatically produce better answers. Duplicates and conflicts can make selection harder.
  • Permission to use an application doesn’t establish permission to read every connected document.
  • Removing a file doesn’t necessarily remove stored copies from every cache or log. The lifecycle needs explicit handling.

Try it

Spend about ten minutes creating a sample document inventory in text. Include a current public outline, a retired outline, and a restricted administrative note. Give each an owner and an audience.

Ask an assistant to propose which materials a learner-facing helper should use and what needs fixing in the inventory. Don’t connect real documents or accounts. Check whether it separates currency from access and asks who owns unresolved decisions. Finish by identifying the permission checks that real software would need to enforce, beyond the written plan.

Keep going

Sources

Share this page