Grounding quality is a content problem, not a model problem

A grounded assistant that answers from a superseded policy has not hallucinated. It has retrieved a document that exists, looks current, and is wrong. The failure is upstream of the model entirely.

Four content faults that produce bad answers

1. Superseded versions retained alongside current ones, with nothing in the document distinguishing them.
2. Documents that contradict each other because two teams own overlapping guidance.
3. Long composite documents where a single retrieved passage loses the qualifying context around it.
4. Content written for an audience that already knows the assumptions, which the assistant reproduces without them.

Diagnose before tuning

When an answer is wrong, retrieve the source it cited. In most cases the source itself is the problem, and no amount of prompt engineering will improve it. Tuning retrieval to compensate for bad content produces a system that is fragile and hard to reason about.

Build an evaluation set early

Twenty to fifty realistic questions with agreed correct answers, run after every change to grounding sources or configuration. This converts quality from anecdote into a number that can be tracked, and it catches regressions when content is edited by people who have never heard of the assistant.

Curate a smaller corpus

A curated set of a few hundred maintained documents outperforms tens of thousands of unreviewed ones, and it can be governed. Expand the corpus when the evaluation set shows the current one cannot answer the questions people are actually asking.

Chunking and structure matter more than they should

Retrieval works on passages. Documents with clear headings, self-contained sections and explicit scope statements retrieve well; wall-of-text documents with implicit context do not. Improving document structure improves answer quality measurably, and it improves the documents for human readers at the same time.