A useful answer starts with a usable source.
- 01InventoryName the source, version, owner, and access.
- 02ResolveSettle conflicts and preserve the context.
- 03Check coverageMap real questions to supported answers.
An AI knowledge base checklist helps you choose which sources a question-answer tool should use. Start with the questions people need answered, then check each document’s owner, current version and intended audience. Resolve conflicts before import, keep context with each passage, and test whether the source set can support useful answers.
This guide concerns ordinary internal reference work, such as studio booking. The inventory and examples are proposed tools, not live test results. Your team’s owners decide which sources may serve the intended readers.
Start with the questions people need answered
Write real task questions before gathering files. “How do I book the studio?” gives you a clear scope. Include a question the system should decline because the source set has no answer.
Microsoft’s RAG preparation guide links source selection to the questions a tool must answer. It also checks file formats and structure. RAG retrieves source material for a generated answer; this checklist prepares that material.
Create a coverage map with a row for each question. Name the source and section that should answer it. If no source supports the answer, mark the gap. Adding more copies of a related file does not fill a missing step in the process.
Build a source inventory with a named owner
Give each document a stable ID and a canonical source: the location your team keeps up to date. A downloaded copy’s filename does not prove it is current. Record where to check for a change.
| Inventory field | What to record | Why the next person needs it |
|---|---|---|
| Document ID and title | A stable reference and plain title | Finds the same source after a rename |
| Canonical source | The maintained location | Avoids treating an old export as the master |
| Version or date | The version used in this import | Connects an answer to the source checked |
| Document owner | The role or person who can resolve a content question | Gives conflicts a destination |
| Intended questions | The tasks this document supports | Shows why the file belongs in the set |
| Audience and access | The approved readers and source limits | Prevents an import from silently broadening use |
| Review trigger | A process change or agreed check date | Makes upkeep part of the workflow |
Ask the document owner to confirm the source. A recent upload may describe an old process. Keep “date uploaded,” “last edited” and “effective version” distinct if your system shows them.
Place unowned files on hold. The person who uploaded a file may not be able to approve it. Give each active source an owner who can resolve a wrong answer.
Resolve conflicting versions before import
Imagine three fictional studio files. Guide v2 says to use the shared calendar. An old PDF says to email the coordinator. A planning note suggests Friday walk-ins. We invented these statements for the example.
| Source | Conflict or gap | Proposed disposition after owner review |
|---|---|---|
| Studio guide v2 | Owner confirms this is the current booking process | Active: use the shared-calendar instructions |
| Studio guide v1 export | Describes the older email route | Archive: keep for history, exclude from the current-answer set |
| Friday planning note | Proposal has no recorded approval | Hold: do not present it as an available option |
Keep a conflict log that names the competing passages and the owner’s answer. The newest filename does not settle the conflict on its own. In this example, the owner confirms v2. If that confirmation is missing, the correct state is unresolved.
Keep the old source under your team’s archive rules. Mark its version clearly and check that the tool no longer retrieves it as current guidance. Archiving a file and excluding it from current answers are separate steps.
Preserve the context around each answer
Review the text after import. A file may look clear on screen while its extracted text loses headings or table labels. Sample each format, including tables and pages with multiple columns where present.
Anthropic’s Contextual Retrieval explanation describes how a passage can lose context when split from its document. Keep the subject and scope with each passage. This worksheet makes no claim to reproduce the vendor’s test results.
For example, “Book two days ahead” is incomplete if it applies only to the large studio. Preserve the heading “Large studio bookings” with that rule. A small-studio question should not inherit a limit from another room simply because the words sound relevant.
Keep table headings with their values. If a row lists a room, a booking window and a contact, check that the extracted text still shows which value belongs to which field. Fix the source or import settings before asking the model to guess the layout.
Check the intended audience and answer coverage
Ask the responsible owner to confirm that each selected source can serve the intended readers. Copying a private file into a shared answer tool changes how people may reach its contents. If the approved audience is unclear, hold that source and use the team’s existing access process to resolve it.
Test with an account intended for the actual reader role. Check that permitted source links open and that the workflow does not expose a source outside that role’s allowed set. This is a task check, not a full security assessment or a guarantee about every possible query.
Then run the coverage map. Ask the ordinary booking question, a question about the old email route and the unapproved Friday option. The expected answers should use the current guide, explain the obsolete route only if the task permits history, and avoid presenting the proposal as approved.
Use the output review checklist to check the answer and its cited passage. Record whether the source was missing, the wrong passage was retrieved, or the draft misread the right passage. Those problems need different fixes.
Before wider use, use the AI pilot scorecard to define a bounded trial and its stop rules. The downloadable inventory includes synthetic rows plus a template row. Keep it with the coverage map, and recheck affected questions when a source changes or is withdrawn.
No comments yet