AI app guides
Evaluate an app's knowledge answers beyond whether they sound right
Check retrieval, support for the answer and coverage of the actual question as separate parts of the journey.
Zocri · 2 min read · Sources checked
Start with questions people actually ask
Microsoft's RAG evaluation guidance separates retrieval quality from answer groundedness, relevance and completeness. Use those distinctions to create a practical review. For an internal service handbook, collect questions about hours, eligibility and ordinary exceptions. Write the expected evidence and answer before testing the system, so a plausible response does not redefine what counts as correct.
Include questions whose answers are absent from the approved documents. The desired behavior may be an honest statement that the information is unavailable, followed by a useful contact route. Do not reward confident completion when the source collection cannot support the answer.
Inspect what the app retrieved
If an answer is wrong, first check whether the relevant material was found. The system may have selected an obsolete policy or a page with a similar title. Record the source version and audience, because an accurate answer from the wrong policy can still mislead the user.
Use a few near-matching documents in the test set, such as different rules for two locations. Ask the same question with the location specified and omitted. Decide when the app should ask a clarifying question instead of choosing a branch's rule without telling the reader.
Judge support and usefulness separately
For each answer, ask whether its claims follow from the retrieved material and whether it addresses the actual question. An answer can stay close to the source while omitting the exception the user asked about. Another can be helpful in tone while inventing a condition that is not documented.
Inspect citations at the statement level. A link to a broad handbook is weaker evidence than the relevant section, but even a precise link must support the accompanying claim. Keep a place for reviewer notes when the source itself is ambiguous or internally inconsistent.
Maintain the evidence as documents change
Assign someone to approve source updates and remove or label obsolete material. Re-run the important questions after a policy change and confirm that old answers are no longer presented as current. Check access too, so a user cannot obtain restricted content through a generated answer.
During a limited pilot, collect unresolved questions and review them before adding new material or changing prompts. Report the evaluation's scope, source versions and remaining gaps. A useful knowledge feature is one that can locate appropriate evidence, answer within it and recognize when the available information does not justify a confident response.