An AI knowledge assistant gives the wrong answer. There are at least two different questions to investigate: did the right evidence reach the model, and did the model use it correctly?
Combining those into one “answer quality” score makes it harder to decide what to fix.
Define retrieval and generation failures separately
A retrieval failure occurs when the needed eligible evidence is missing from the supplied context. A generation failure occurs when sufficient evidence is present but the answer misstates, ignores, or extends it incorrectly.
There are also questions with no supporting answer in the collection. For those, the correct behaviour may be to acknowledge the gap rather than manufacture a response.
Build a small evaluation set with known sources
Use fictional or approved documents and record the expected supporting passages before running the assistant.
| Question type | Retrieval expectation | Answer expectation |
|---|---|---|
| One documented fact | Relevant eligible passage included | State the fact accurately |
| Several supporting passages | Necessary evidence included | Combine only supported claims |
| Obsolete policy | Current applicable version selected | Respect version and scope |
| No documented answer | No false evidence requirement | Acknowledge the information gap |
Keep evidence labels independent of the assistant's own explanation. A generated justification should not decide whether its supporting passage was actually sufficient.
Inspect context before adjusting the model
For each failed answer, save the question, retrieval filters, passage references, source versions, and final answer under the appropriate access controls.
If a key passage is absent, inspect parsing, indexing, metadata filters, and retrieval ranking. If the passage is present, inspect answer instructions, context boundaries, conflicting evidence, and claim formation.
The PDF extraction guide covers a failure even earlier in the pipeline. The policy-version guide covers an example where a semantically relevant source is still inapplicable.
Report metrics with clear denominators
If you measure how often a required passage is retrieved, define the eligible questions and passage labels. If you measure supported claims, define what counts as a claim and how reviewers judge support.
Avoid presenting a small, hand-selected test set as a general accuracy benchmark. Record its size, limitations, and review method. Keep permission-sensitive cases separate so strong answers do not hide access failures.
For release decisions, preserve the same set across versions and add regressions discovered later. A changing test set can make an apparent improvement hard to interpret.
Choose the fix that matches the failure
Increasing model capability may help some generation problems while leaving missing evidence untouched. Expanding retrieval can add the needed passage but also introduce more conflicting material.
Make one targeted change and rerun the relevant cases. Use the source-citation article to review whether the answer's individual claims remain supported.
Knowledge evaluation questions
Can retrieval pass while the answer fails?
Yes. Relevant evidence can be supplied and then misused.
Should unanswered questions count as failures automatically?
Not when the collection genuinely lacks the answer and the assistant handles that honestly.
What should a review deliver?
A classified failure set and the evidence behind each classification. Use it to scope the next engineering change.