Build an internal knowledge base your team can maintain

September 5, 2026 · Ask Your Docs

An internal knowledge base is a maintained collection of reference material for an authorized team. Its value depends on whether someone can find the applicable rule, identify its source and know what to do when it is incomplete. Start with a bounded workflow and make ownership, access and updates part of the first release.

This guide uses a fictional sales policy to demonstrate a maintenance process. The checks below are expected outcomes for an exercise, not results from a customer deployment or a completed AI benchmark.

Define the decision the library should support

Choose a recurring decision with a clear boundary. For a sales team, that might be preparing an offer for a new annual subscription. The library should explain which discount rules apply, which approvals are required and where to record the request. It should also make clear that finding an approver’s name does not prove approval was given.

Write a handful of real questions before collecting documents. Include a straightforward request, a boundary case, an exception and a question the sources cannot answer. For the fictional Cedar exercise, those are a 10% discount, a 12% discount, permission to reuse a customer quotation and the refund policy. The final question deliberately falls outside the supplied policy.

Agree on who will use the answers and what action still requires a person. A reference library can explain a process without being the system that approves an offer. Keep that distinction visible in the source and in any assistant instructions.

Inventory sources before uploading a department folder

Build a source register that connects each question to its authoritative document. The starter kit includes an editable register. Use it to identify duplicates, uncertain ownership and references that only exist in someone’s chat history. A filename containing “final” is not enough to decide which version governs a decision.

Retain original spreadsheets and presentations when their meaning depends on formulas or visuals. A text reference can explain which workbook or deck to consult; it should not silently replace calculations or diagrams with an incomplete extract.

Assign ownership to rules, not just folders

Separate three responsibilities: deciding the rule, maintaining its written source and administering access to the system. One person may hold several responsibilities in a small team, but each responsibility still needs to be explicit. A librarian should not rewrite a commercial exception merely to resolve a confusing search result.

In the Cedar example, a discount above 10% requires both sales and finance approval. The owner of the discount policy maintains that rule. The escalation directory identifies the fictional people associated with those roles. Changing a person in the directory should not require copying the threshold rule into another document.

Record a replacement owner when someone changes roles or leaves the team. If no one can confirm whether a source remains valid, label the uncertainty and decide whether to remove it from active guidance. An apparently current document with no accountable reviewer can produce confident but unreviewed answers.

Verify access independently of content quality

Describe the audience required for each source, then check whether the chosen storage and assistant setup enforces that boundary. A folder name, metadata field or instruction saying “internal only” is a description, not an access control. Do not infer that a document’s original permissions follow it into a second system.

Use a harmless, clearly fictional test document when checking access. With authorized test accounts, verify that an intended reader can retrieve it and an excluded reader cannot. Repeat the check through each route people will use: direct browsing, search, shared links and an assistant connection. Test the content returned, not merely whether a navigation item is hidden.

As of September 2026, this guide does not claim that every platform provides the same team controls. Check the actual account and configuration. For Ask Your Docs, read the document processing explanation before introducing internal material. If a required boundary cannot be enforced, keep that material outside the proposed collection.

Make an update a small release with evidence

An update has four parts: approve the changed rule, replace the active reference, verify retrieval and record the result. Treat each part separately. Saving a revised local file does not establish that the library or an external assistant can see it.

Practice with the Cedar Markdown example. In a copy of discount-policy.md, change commercial quote validity from 14 to 7 calendar days and increase the version from 1.0 to 1.1. This is an exercise change, not a recommendation for a real commercial policy. Replace the active source using your chosen system’s documented workflow.

Ask the validity question in a fresh conversation. The expected answer is now 7 calendar days with evidence from version 1.1. If it still says 14, inspect the returned source and connected collection. Preserve history in an appropriate archive, but avoid leaving indistinguishable old and new rules in the active reference set.

Review the answer and the evidence separately

A correct sentence can still have inadequate evidence. Asking who approves a 12% discount should require the threshold rule and the role directory. A citation to the directory alone supports a person’s role, but does not establish the discount threshold. Check both parts before treating the answer as usable guidance.

Also test a question that combines two independent policies. Discount approval does not authorize an attributed customer quotation: the fictional quotation policy requires written permission and legal approval. This catches an answer that wrongly treats one approval as covering the entire sales workflow.

Keep expected answers outside the collection under test. The example download separates four source files from the evaluator’s CSV for this reason. Record the actual answer, cited file and version, missing evidence and reviewer decision. A failure should produce a specific action: repair the source, fix access, replace an obsolete copy or escalate an unsupported question.

Choose a review cadence from the consequences of change

Review sources when their underlying rules change, not only when a calendar reminder arrives. Useful triggers include a new product definition, a changed approval threshold, a departed owner or a question that exposes conflicting instructions. Add a scheduled check as a backstop, with an interval chosen by the source owner.

Measure maintenance with observable records. Count questions with missing sources, answers using obsolete versions and changes awaiting review. Keep the denominator and review period visible. For example, “three of twelve reviewed questions lacked sufficient evidence” is interpretable; a standalone “75% accurate” label hides what was tested.

Do not use page views alone to conclude that the library is helping. A heavily visited page may be useful or confusing. Pair usage with reviewed questions and owner feedback. Expand the collection when you can identify an unanswered task and the maintained source needed to resolve it.

Connect an assistant after the reference process works

You can begin with a readable library and add AI retrieval later. OpenAI’s ChatGPT Projects documentation and Anthropic’s Claude project instructions describe native reference workflows. Confirm their current account requirements and sharing settings before moving team documents.

If you use a connected document library, repeat the same source, access and update checks through the connection. The underlying maintenance responsibilities remain. Upload a file, organize the references, connect the intended assistant and verify an answer with sources. A successful setup screen alone is not proof that the intended source is available.

For a portable starting corpus, follow the Markdown library guide. For the broader role of retrieval across different file types, read the AI knowledge base guide. Keep this internal library’s operational record focused on its owners, audience, current sources and unresolved questions.

Frequently asked questions