Fetching more passages often makes answers worse, not better. What we do before the model ever sees a document.
A thought to take with youChoose fewer, better passages. Filtering and ranking matter more than how many results you pass to the model.
Why more passages can hurt
When answers are weak, the tempting fix is to retrieve more: ten passages instead of five, twenty instead of ten. In practice, the extra passages are often similar in wording but wrong for the question: an old version of a policy, a draft, or a document the person should not see.
The model then has to decide between conflicting sources, and it does not always pick the right one.
What we do instead
We filter first on facts we already know: date, status, owner, and who is allowed to read it. Then we combine keyword search with meaning-based search, and re-rank the short list so the most relevant passages come first.
Only a handful of passages reach the model, and each one carries its source so the answer can point back to it.