Knowledge base examples: three complete business libraries
A useful knowledge base connects a recurring question to maintained sources and a clear decision owner. These three examples show how that works for a marketing campaign, a sales policy and workshop operations. Each is complete for its small exercise, with source files and separate reviewer questions.
The companies, people and policies are fictional. The download contains 11 source files and 14 authored evaluation questions. It does not contain customer results, real approvals or a completed benchmark of ChatGPT or Claude.
Choose one library, upload only its source files and keep the evaluation sheet outside the collection.
Download all three example libraries →Compare the three knowledge base examples
Each collection supports one bounded workflow. Marketing focuses on what a campaign document actually authorizes. Sales combines a threshold policy with role ownership. Operations applies a capacity constraint before communicating a room change. Keeping the scopes separate prevents an assistant from treating unrelated companies or policies as one business.
| Library | Source files | Key question | Deliberate challenge |
|---|---|---|---|
| Harbor marketing | Word brief, PowerPoint deck, Markdown index | What still needs approval? | Budget status differs between files. |
| Cedar sales | Four linked Markdown files | Who approves a 12% discount? | Two roles and a separate quotation exception. |
| Juniper operations | Four linked Markdown files | Can the workshop move rooms? | The proposed room is too small. |
These are internal reference examples. A customer help center has a different audience and navigation problem. Help Scout’s examples illustrate public help-center organization; use the downloadable collections here to inspect the source and answer relationship behind an internal workflow.
Marketing example: campaign approvals across Word and slides
The Harbor library contains an actual DOCX brief and a four-slide PPTX deck. Both describe an October 12, 2026 campaign owned by Priya Shah. The files give the same $18,000 budget amount, but the brief calls it approved while the deck describes it as planned. The library index does not declare either source to supersede the other.
Ask, “Is this campaign ready to run paid advertisements?” The expected answer should preserve the finance approval still required for paid ads. The word “approved” elsewhere in the brief cannot stand in for that specific approval. Ask for both the supporting passage and the activity it applies to.
Then ask about adding a customer quotation. Legal review is a separate condition, and the deck also requires written customer permission. This checks whether the assistant carries an exception into the answer. Read the document analysis guide for a field-by-field comparison, including how to record the unresolved budget status.
Sales example: connect a rule to its actual owner
The Cedar library separates the discount policy, role directory and customer quotation policy, with an index connecting them. A discount of 10% or less requires the sales lead; above 10% it requires both sales and finance. The directory identifies fictional role owners Alex Chen and Morgan Lee.
Ask, “Who must approve a 12% discount on a new annual subscription?” A complete answer needs the threshold rule and the directory. A citation naming Morgan alone does not establish why finance is required. Conversely, a threshold citation alone does not establish which person currently holds that role.
The policy limits its scope to new annual subscriptions and does not define refunds. The question “What is the refund policy?” should therefore produce a missing-information answer. Do not import an assumption from the marketing or operations example. Follow the Markdown knowledge base guide for the full four-file corpus and an update exercise.
Operations example: apply a constraint before taking action
The Juniper workshop is scheduled for September 21, 2026, with check-in at 09:00 ET and a 09:30 ET start. Fourteen participants are registered for Atlas, an 18-seat room. Birch appears as a possible alternative but has only 12 seats; its availability is not confirmed.
Ask, “Can we move everyone to Birch and send the new instructions?” A source-backed answer should identify the capacity mismatch and the missing approvals. Sam Rivera must approve the revised event plan and Taylor Brooks must confirm the room booking. Casey Lin sends participant instructions only after both responses are recorded and the plan fits the room.
The procedure does not authorize silently excluding two participants, adding chairs or splitting the workshop. Those would change the plan. This example tests whether an answer preserves an operational condition instead of turning a named alternative into a confirmed solution. The sources also deliberately omit a catering cancellation fee.
Use the download without leaking the answer key
Extract the ZIP and choose one of the three library folders. Upload the files inside that folder’s sources directory individually if your system requires it. A downloadable ZIP is packaging; it does not imply that an assistant accepts ZIP uploads or automatically follows links between the documents inside.
Keep evaluation.csv and the top-level README outside the collection under test. The evaluator contains expected answers, so uploading it could make a correct-looking response simply retrieve the answer key. Fill in its blank actual-answer, actual-source and review-result columns after each question.
For the marketing example, inspect the original slides if a question depends on images or speaker notes. For the Markdown collections, upload every required source rather than assuming a relative link will fetch it. The ChatGPT setup guide and Claude setup guide provide routes to check in your account.
Evaluate conditions, sources and missing information
Review the answer and its evidence separately. First check whether the conclusion is correct for the question. Then open each cited passage and decide whether it supports the whole claim. Finally, check for omitted conditions that would change what the reader should do next.
A useful evaluation set includes a direct fact, a boundary case, a question requiring two sources and a missing fact. The Cedar 10% boundary and the Juniper room-size mismatch are examples. Repeat the question in a fresh conversation after correcting a source, so earlier conversation text does not conceal whether the updated reference was retrieved.
Record failures in plain language: “finance condition omitted” is more actionable than “bad answer.” Keep an unresolved result open until its cause is understood. These examples provide review material; they do not assign a universal accuracy target or guarantee an assistant will always find the relevant passage.
Adapt one example into a maintained team library
Choose the closest workflow and replace the fictional rules with approved sources. Assign a source owner, intended audience, version and review trigger. Preserve uncertainty when nobody can establish which document is current. A clean-looking library with unresolved rules remains unresolved.
Check access through the actual system people will use. A heading saying “internal” or a metadata audience field does not enforce permissions. Begin with harmless examples when testing who can retrieve content. Read the document processing explanation before uploading private material to Ask Your Docs.
For a public help center, also design the path from a common question to the relevant article and a next step when the answer is missing. Help Scout’s article-writing guidance is a useful companion. For internal ownership and updates, use the team maintenance guide and its source register.
Frequently asked questions
Can you give me some examples of knowledge-based systems?
A rule-based decision system, a company reference library and a support knowledge base are examples of systems that organize knowledge for decisions. Their capabilities differ: a library provides references, while a rule engine may execute explicitly defined rules. Adding an AI answer interface does not itself grant approval authority.
What is a knowledge base?
A knowledge base is an organized collection of reference material for a defined audience and purpose. It can contain maintained articles, original files, links and supporting records. Useful answers depend on source quality, access, current versions and a way to resolve missing or conflicting information.
Can you give me some examples of knowledge work?
Preparing a campaign brief, interpreting a discount policy and planning a workshop are examples of knowledge work. Each involves applying information to a situation. A reference library can support the work, while the people responsible still make and record the required decisions.
What is the best knowledge base?
The best fit depends on your audience, tasks, permissions and maintenance process. Test a small representative collection with real questions before choosing. Check whether readers can find sufficient evidence, whether updates become visible and whether excluded readers remain excluded.