RAG gives confidently wrong answers: upgrade the model or fix the data?

Conceptual source: Microsoft Learn’s RAG evaluators. The documentation separates retrieval evaluation from answer quality. The workflow and examples below are independent analysis, not a product introduction or real deployment report.

Imagine asking an internal assistant about purchasing equipment. Its fluent answer includes a link, but points to an outdated procedure. This is not necessarily just model hallucination. The model may have faithfully summarized the wrong document.

One wrong answer, several possible causes

RAG usually retrieves document passages for the model. Failure may precede generation: the correct document was never indexed, the right passage was missed or an obsolete version ranked first. Alternatively, the right text arrived but the answer exceeded its evidence.

I do not consider RAG quality a single score. Separate failure to find evidence from failure to use it. A more accurate model receiving the same incorrect context will not necessarily deliver an acceptable answer.

A useful citation is more than an attractive link

A link does not prove the document supports a claim. Reviewers should locate the relevant passage and understand its date and scope. Faithfulness also does not establish present-day validity: stale material can yield a grounded but unacceptable answer.

This is a proposed framework, not a tool’s guaranteed feature list. Content conflicts need an owner and organizational rule; the model should not silently become the policy authority.

Build the evaluation set from real questions

1. Collect recurring questions without confidential information.
2. Identify expected documents and passages with their owner.
3. Add unanswerable, ambiguous, outdated and access-restricted cases.
4. Record retrieval, citation correctness and answer acceptance separately.

Compare changes against that set. Changing model, instructions, chunking and search together obscures the cause of improvement or regression. One successful demonstration does not represent all employee questions.

Permissions belong in quality evaluation

An accurate answer exposing information to an unauthorized person is not product success. Check how document access is enforced and whether prohibited material leaks through hints or summaries. Logs must not become an unprotected archive of sensitive text either: retain necessary identifiers and metadata with appropriate access.

Decision questions

When should we upgrade the model?

When correct, authorized evidence arrives but the model mishandles it or the required response format, and controlled testing demonstrates an acceptable improvement. Cost and latency remain acceptance criteria.

Is automated evaluation sufficient?

It helps repeat comparisons, but evaluators can be wrong. Borderline cases and sensitive answers need content-owner judgment. A good score is not immunity from error.

Who maintains document freshness?

Decide beforehand. I suggest ownership, review dates and removal procedures for obsolete versions. “The model will notice” does not replace maintenance.

The final measure is an accepted answer to an authorized question, supported by evidence that actually backs the claim. For the performance side, read CPU and GPU bottlenecks.

Source checked: November 2, 2026. Examples are hypothetical and the proposed method is original, not a guarantee of model or system accuracy.