A close match is a starting point
Ask a document search system an unrelated question and it may still return a passage. That is not necessarily a broken search engine. Nearest-neighbor search looks for the closest candidates in the available collection, even when none of them answers the question.
Imagine a handbook that contains only support hours and file-upload instructions. A question about an office address still has a nearest passage. The passage is simply the closest thing available. It does not establish where the office is.
A similarity score describes a relationship between representations. It is not a probability that a passage proves a claim. Treating those two ideas as interchangeable makes an assistant sound certain when its sources are silent.
Separate finding from answering
Retrieval collects candidate passages. Answering asks a different question: do these passages contain enough information to support the requested claim? Keep that second decision explicit.
Keyword search can preserve exact names and amounts. Vector search can find different wording for a similar idea. Combining them can improve the candidate set, but neither mechanism removes the need to inspect the evidence.
For example, a passage saying “Support is available Monday to Friday” can support a question about support days. It cannot support an exact response time, even though both questions concern support.
Make missing evidence a useful outcome
When the available sources do not support an answer, say so. A clear insufficient-evidence result helps the reader decide whether to upload a missing source, narrow the question, or inspect the original material.
In ContextPier, search exposes passages while grounded questions can return an insufficient_evidence outcome. These are different views of the same knowledge: one helps you explore; the other needs enough support to answer.
A practical test collection should include answerable questions, unrelated questions, and questions that require a detail the source leaves out. A system that can decline the last two is easier to trust than one that always produces a fluent answer.