Knowledge base starter kit for your existing files
Download an editable knowledge base template pack for organizing sources, assigning owners and checking AI answers. It includes Markdown reference-page templates, CSV registers you can open in a spreadsheet, and fictional Word and Excel examples. No account is required to download the files.
One ZIP with nine files. Keep the originals and adapt a copy.
Download the starter kit →What is in the download?
| File | Purpose | Use it when |
|---|---|---|
| Source register | Track owner, version, dates and access scope. | Collecting or reviewing sources. |
| Reference-page template | Document facts, exceptions and missing information. | A recurring answer has no maintained source. |
| Assistant instructions | Request source evidence and explicit uncertainty. | Setting up a project or library workflow. |
| Evaluation questions | Compare actual answers with known results. | Checking the fictional examples. |
| Update log | Record changes and affected questions. | A source is replaced or withdrawn. |
| README and three example files | Instructions plus a DOCX policy, XLSX workbook and CSV table. | Practicing before using internal material. |
The files are ordinary Markdown, CSV, DOCX and XLSX. You can inspect and edit them using tools that support those formats. This pack is a starting structure, not a hosted database or an automatic synchronization service. The templates do not enforce the ownership and access rules you write into them.
Start with one recurring question
Choose a question people currently answer by finding a document: whether a campaign can launch, which product claim is approved, or who owns a handoff. Write the question and its source before selecting more files. A small collection with a clear purpose is easier to maintain and test than a folder of vaguely related documents.
In the source register, replace the fictional rows with your own material. Assign an owner who can confirm whether each source is current. Add a version and review date. Describe the intended audience in the access field, then apply that decision in the system that actually stores or shares the file.
If two sources disagree, record the conflict and resolve it with the owner. Do not silently choose one because its filename says final. An AI answer cannot repair an unresolved policy decision simply by citing a passage.
Use the page template for missing source material
The Markdown page template separates purpose, authoritative information, exceptions, examples and information not covered. These fields help a human reader understand the rule before an AI system retrieves it. Keep facts in complete sentences and preserve important dates, units and conditions.
Use a new page when an answer exists only in someone’s head or a scattered conversation. Do not rewrite a well-maintained original merely to populate the template. Link to the authoritative source and identify what your summary adds. Every duplicate introduces another place that can become stale.
Label examples clearly. A fictional scenario is useful for training and testing, but it must not look like customer evidence. The supplied Harbor campaign and sales workbook are separate exercises; the documents do not establish that the sales revenue funds the campaign.
Keep evaluation answers outside the source collection
The evaluation CSV includes expected answers for the fictional files. Keep it open beside your test, but do not upload it into the reference collection being evaluated. Otherwise the assistant can repeat an answer from the checklist without finding the original policy or table.
Upload the source document, ask the question, then record the actual answer and supporting reference in the empty columns. Verify both. A correct conclusion with an unrelated citation is weak evidence. A missing fact should produce an admission that the source does not answer the question, rather than an invented value.
The Word exercise tests a general rule, current status and an exception. The Excel exercise tests currency, date and status filters. The workbook includes stored check values for human review; require the underlying records and method when evaluating calculation behavior.
Set up the same sources in your chosen assistant
Native projects are one way to try the examples. OpenAI documents reference files and instructions in ChatGPT Projects; Anthropic describes reusable Claude project knowledge. Check current limits and workspace controls before uploading internal material.
Use the ChatGPT guide or Claude guide for the setup sequence. If you use Ask Your Docs, organize the sources and follow Connect your AI, then repeat the source checks through that connection.
This template does not guarantee identical behavior across assistants. Record the route, account context, date and source versions for each evaluation. The same prompt can only support a useful comparison when the intended sources and access are also comparable.
Record updates and rerun affected questions
Use the update log when a source changes. Record the old and new versions, the change, the person responsible and the questions that need another check. Replace or remove obsolete active references according to your organization’s rules. Historical records should be clearly separated from current guidance.
For practice, change the fictional campaign launch date from October 12 to October 19, 2026 in a version 1.1 copy. Update the active source and start a fresh conversation. Ask for the date and source version. Repeating a previous chat answer does not prove the replacement was retrieved.
The broader AI knowledge base guide explains how to grow the collection after these checks work. Expand around actual questions, keep owners accountable for sources and measure useful supported answers rather than file count.
Can I use this without Ask Your Docs?
Yes. The files are directly downloadable and do not require an Ask Your Docs account. They provide an editable structure you can adapt to your chosen workflow. The pack contains fictional examples and suggested operating instructions, not legal advice, access enforcement or a claim that a particular AI system has passed the tests.