Build a knowledge foundation for enterprise AI.

Begin with the questions a team needs to answer. Identify the authoritative sources for those questions and the information that should remain outside the first release. A focused library is easier to inspect and evaluate than an indiscriminate import.

Distinguish approved guidance from drafts, historical material and personal working notes. Where several documents disagree, assign an owner to resolve the conflict or define how the system should present the uncertainty.

Document preparation can remove the very context a reader needs. Headings, table labels, dates, version identifiers and links between sections can change the meaning of a passage. Inspect representative material after extraction and indexing.

Plan for the actual content formats, including tables and scanned documents. Test Arabic, English and mixed-language material with appropriate reviewers. Keep a reference back to the original source so that the answer can be checked.

For a scanned table, compare the extracted result with the original page. A shifted column or missing unit can change the meaning of an answer. Include examples with the layouts and document quality your users actually rely on.

Define who can retrieve each source and how that permission is enforced. Test with users who have different access levels, including users who should receive no result. A convenient shared index needs a considered access model.

Assign refresh and retirement rules. Record how a policy update, source deletion or access change propagates into the knowledge experience. Users need a way to recognise a source version and report material that appears incorrect.

Illustrative example: a library contains an approved policy and an older procedure with a different instruction. Simply indexing both leaves the assistant to work with a conflict that the organisation has not resolved.

Ask the source owner which document governs the question, preserve the approval and version context, and decide how historical material should be retrieved. Add the conflict to your evaluation set. A useful result identifies the relevant current source or explains that the available material does not resolve the question.

For each selected source, record its purpose, owner, location, approval status, languages and intended audience. Note the content formats and the questions it is expected to support. This can begin as a small working table; its value comes from the decisions it makes visible.

Include what happens when the source changes. A new version may replace earlier guidance, add a new process or change who can access the material. Define how the knowledge experience learns about that event and who confirms the updated result. Check deletion and withdrawal paths as carefully as the initial import.

Use the register to identify gaps before attempting to solve them through prompting. If no approved source answers a common question, the next action may belong with a business owner. Capture the gap, its audience and its priority. A well-structured knowledge foundation can reveal useful content work even before an assistant is released.

Choose representative questions and identify the passages a reviewer would expect to support a response. Run them through the retrieval workflow and inspect the returned context. Check relevance, permissions, version and whether important headings or table labels survived preparation. Then review the generated answer against that context.

Separate the results so the team can choose the right fix. An absent passage may indicate a source or retrieval problem. A returned passage with a missing qualification may indicate preparation or answer-generation work. A valid answer with an inaccessible source link is still a usability issue that deserves a clear owner.

Microsoft’s RAG design and evaluation guidance provides a technical reference for teams exploring retrieval-augmented generation. Use it alongside questions and review criteria drawn from your own workflow. Preserve the evaluation set as the library grows so that new content can be checked without losing sight of earlier use cases.

If a system retrieves the wrong passage, an eloquent answer does not fix the underlying problem. Review whether representative questions retrieve relevant, permitted and current material before judging the generated response.

Include questions for which the library has no answer. The desired behaviour may be to explain the gap and direct the user to an owner. Keep improving the source set and the evaluation examples together. Learn more about our Aliph Data services and AliphChat.