# AI knowledge base readiness checklist

Use this checklist before you connect an AI assistant to your knowledge base, and again before each expansion to a new team or audience. It goes with the AFFiNE guide [What Is an AI Knowledge Base?](https://affine.pro/blog/ai-knowledge-base).

Mark each item as done, partly done, or not started. Fix the "not started" items in the Sources, Content, and Permissions groups before a pilot.

## Sources

- [ ] 1. One starting use case is written down, for example "answer IT questions for new hires."
- [ ] 2. Every source for that use case is listed: wiki pages, help articles, PDFs, tickets, chat threads, transcripts.
- [ ] 3. Each source has a named owner (a person, not a team).
- [ ] 4. Topics the assistant must not answer are listed, such as pricing exceptions, legal advice, or personal data.

## Content

- [ ] 5. Duplicate pages on the same question are merged, with one canonical page per question.
- [ ] 6. Outdated pages are retired or clearly marked, not left in the index.
- [ ] 7. Section headings name the topic, and the answer is in the first sentence of each section.
- [ ] 8. Sections avoid references like "as mentioned above" that break when a chunk is read alone.
- [ ] 9. Similar pages state their scope in the first sentence (product, version, audience, region).
- [ ] 10. Each page has metadata: owner, last reviewed date, audience, product and version, content type.

## Permissions and data

- [ ] 11. The retrieval layer respects source permissions: no answer uses a page the user cannot open.
- [ ] 12. Internal-only content is excluded from any customer-facing assistant.
- [ ] 13. You know where documents are stored and, separately, where AI processing runs and which provider processes prompts and retrieved passages.
- [ ] 14. Personal data in tickets, transcripts, and screenshots is removed or masked before indexing.

## Testing

- [ ] 15. A golden question set exists, built from real questions, with the expected answer and source for each.
- [ ] 16. The set includes questions the sources do not cover, permission checks, outdated-content traps, and ambiguous questions.
- [ ] 17. Retrieval and answers are scored separately, and every miss gets a failure code.
- [ ] 18. A baseline is recorded, and the set is re-run after every change to content, prompts, or models.

## Ownership and operations

- [ ] 19. One person owns the feedback loop: reviewing flagged answers and turning confirmed failures into new test cases.
- [ ] 20. Review dates are scheduled for each content type, and stale pages are flagged automatically or on a calendar.

---

This is a template. It does not contain test results. Adapt the items to your tools, content, and policies.
