Set a Knowledge Freshness Policy for Voice Answers

A voice assistant knowledge freshness policy decides when information is current enough to say aloud. Uploading a document once does not make its contents permanently true. Opening hours change, services are withdrawn and an old price sheet can remain accessible long after the business stops using it.

The policy should be simple enough for an operator to maintain. Its purpose is to prevent outdated information from becoming a confident spoken promise, while giving callers a useful alternative when the latest answer is unavailable.

Classify information by how it changes

Group knowledge into stable background information, periodically reviewed information and rapidly changing state. A business description may remain stable for months. Seasonal hours need a review trigger. Live availability may require a supported lookup rather than a static document.

Do not give every source the same expiry period merely for convenience. Ask the owner what event makes the information unreliable. A catalogue revision, holiday schedule or supplier notice may be more meaningful than an arbitrary monthly reminder.

LiveKit's external data guide describes connecting an agent to retrieved information and other systems. The retrieval mechanism still needs an authoritative source and a business rule for freshness.

Create a source register

For each approved source, record its owner, effective date, last review date, intended audience and replacement rule. Include where the authoritative version lives. A filename containing “final” is not a reliable version policy.

A hypothetical service business could maintain one approved hours document with a named owner and a holiday review trigger. When a new version takes effect, the old one is removed from the assistant's active knowledge selection or otherwise made unambiguously obsolete through the supported process.

Burki's production-safe learning guide provides a useful review principle: changes should be inspected and accepted, not assumed to improve the assistant automatically.

Define what happens when sources conflict

Decide which source wins before a caller discovers the conflict. A current operating policy may take precedence over an older brochure. A live supported lookup may govern current availability, while the knowledge document explains the general process.

If the system cannot establish which source applies, the assistant should avoid guessing. Write a short fallback that states the uncertainty and offers a supported next step. Do not let the model choose a convenient answer simply because both documents contain plausible wording.

Also distinguish information that is unavailable from information that is negative. An expired availability source does not mean no appointments, products or services exist; it means the assistant cannot reliably answer from that source.

Test the changed facts directly

After an update, ask the question in several natural ways. Include a question that refers to the old fact and a request for an exception. Check that the assistant uses the current information without inventing a broader policy.

For the hypothetical hours change, test “When do you close?”, “Are you still open until the old time?” and “Can I come after closing?” The final question checks whether the assistant turns an updated schedule into an unsupported exception.

Keep a short record of the knowledge revision and the observed answers. If the answer requires a tool, inspect its result as well as the spoken response.

Keep withdrawn facts in a regression case so an old answer returning after a later update becomes visible.

Assign maintenance to a real owner

A freshness policy fails when everyone assumes someone else updates the documents. Name the primary owner, a backup and the events that trigger review. Include urgent withdrawal: if a source is discovered to be wrong, the team needs a way to stop relying on it while a replacement is prepared.

Avoid measuring success by the number of uploaded pages. Measure whether the important questions receive current, supported answers and whether uncertainty is handled honestly.

Use Burki's testing guide to keep a small question set beside the source register. Start with the five facts most likely to change and assign an owner and review trigger to each.