Preparing Institutional Content for an AI Knowledge Assistant

Start with a content inventory

An institutional knowledge assistant needs a defined collection of approved information. Begin with the tasks it will support and inventory the sources relevant to those tasks. Include owner, audience, publication date, review date, format, access restrictions, and authoritative location. A shared folder is a storage location; it is not evidence that every file inside should be available to the assistant.

Classify material as approved, needs review, obsolete, restricted, or outside scope. The initial collection should be small enough for an owner to inspect. An illustrative student-services pilot might begin with current contact details, a service handbook, and approved process pages. Draft meeting notes and archived versions would remain outside the collection until someone establishes their authority and relevance.

Resolve conflicts before indexing

Identify sources that disagree about dates, amounts, steps, eligibility conditions, or office responsibility. Ask the responsible team to resolve the conflict rather than expecting a model to guess which wording is correct. Record precedence rules where several valid sources cover different audiences or periods. Preserve those distinctions in the content metadata.

An illustrative institution may publish separate instructions for continuing students and new applicants. Both can be current, but a general answer assembled from both may be wrong for either group. The assistant should ask for the missing context or show the relevant distinction. Organising sources by applicability reduces ambiguity before retrieval and makes later review more straightforward.

Inspect what the system can actually read

Check the extracted representation of PDFs, tables, equations, slide decks, and scanned pages. A file that opens correctly for a person can produce incomplete or disordered text in an ingestion process. Inspect headings, table relationships, footnotes, and exceptions. Where a resource cannot be extracted reliably, supply an approved alternative or exclude it from the supported scope.

Create a content check using a few questions whose answers depend on awkward formatting. A fee table should preserve the relationship between programme, intake, and amount; a policy footnote may contain an important exception. Ask the assistant to show the source passage used. A correct-looking answer without a reliable evidence trail can conceal an ingestion problem that will affect less familiar questions.

Keep permissions attached to the content

Specify which audiences can retrieve each collection and how the system verifies that access. Public FAQs, enrolled-student materials, staff-only procedures, and confidential assessment resources should have distinct handling. If the assistant creates a separate index, document how access changes and deletions propagate to it. Test retrieval using different roles rather than assuming the indexing tool inherits institutional permissions.

Treat retrieved text as data. A document can contain instructions that attempt to change assistant behaviour, including content copied from outside the institution. OWASP's prompt-injection guidance supports separating untrusted content from system authority. Access checks, permitted actions, and external destinations should be enforced by the application, with the source text unable to grant itself extra permissions.

Create a repeatable update process

Name who approves additions, corrections, and removals. Define the trigger for refreshing content and the evidence that a change reached the searchable collection. Record a version identifier for the content set so a reported answer can be investigated against the information available at that time. Recheck questions affected by a changed policy or source structure.

An illustrative admissions update might change a document requirement for the next intake. The owner approves the new source, the system replaces or scopes the old version, and a small test confirms the right answer for each intake. This is more dependable than adding the new page while leaving an undifferentiated older copy available. Retention and archiving decisions remain with the institution's established process.

Use a readiness review before development expands

A content readiness review should reveal missing authority, poor extraction, unclear audience boundaries, and unsupported questions. Those findings are useful project outputs. They help the institution decide whether to improve sources, narrow the assistant's task, or plan additional review capacity. More documents do not automatically improve an answer if they add contradictions and irrelevant material.

• Choose sources from the assistant's intended tasks and identify a responsible owner.
• Check authority, audience, dates, permissions, extraction quality, and conflicting versions.
• Test source-linked answers and clear responses when information is missing.
• Agree how approved changes, removals, and access updates reach the live assistant.