Answers and evidence

Every tool returns the same two things: a one-line headline for a human to read, and the full record as JSON. The headline is a summary and its wording can change; the JSON is the part to rely on, and it carries the ids that let an assistant chain one call into the next.

Asking a question

ask_question runs the same pipeline as asking from the home page. It returns the answer along with everything behind it: how confident ResponseHub is, why, and which material it drew on.

Answering is not instant. If the pipeline is still running, the result says the question is still processing and gives its id — the assistant fetches the finished answer with get_question. A well-behaved assistant does this for you; you may see it pause and check back.

Confidence and relevance

Both are scored 0.0–1.0 and reported as a percentage and a word, so nothing has to interpret the number. The bands are the same ones the web UI uses:

Words	Score
Excellent	80–100%
Very good	60–79%
Good	40–59%
Ok	20–39%
Low	0–19%

Confidence describes the answer as a whole. Relevance describes a single reference — how well that piece of material matched the question. The full explanations live in confidence and relevancy.

An answer also carries a rationale for its confidence, and a verdict on whether there was enough context to answer at all. When there was not, ResponseHub says so rather than guessing — the result explains what was missing instead of offering a low-confidence answer as though it were fine.

The references behind an answer

An answer result carries every reference the search step surfaced, not only the ones quoted — used ones first, then in relevance order. Each reference says:

Whether the answer actually used it.
Whether it is a knowledge base item or a source.
Its title, and an excerpt of the matched text.
A web link into ResponseHub, so you can open the underlying record.
The character positions of the exact quoted span within the excerpt, for assistants that highlight it.

The result also counts them separately: how many references were found, and how many were used. The used count is the meaningful one — the rest is what search considered and the answer set aside.

The answer card

Some assistants can render a ResponseHub answer card instead of plain text: the answer, a confidence bar, the rationale, and the reference cards with the quoted passage highlighted and unused references tucked behind a toggle.

This is a display layer, nothing more. Assistants that cannot render it show the text result, and the text result is what the model reads in both cases — the card never changes the answer.

Source excerpts are shown exactly as ResponseHub extracted them, which for a scanned PDF can include headings, bold markers and table pipes from the original layout. That is the raw extracted text, not a formatting error.

Lists and limits

List and search tools cap how much they return. Ask for more than the maximum and the request is not rejected — it is clamped to the maximum, and the result reports both what was asked for and what was applied, so nothing quietly returns a partial set that looks complete.

Tool	Default	Maximum
`search_knowledge_base`	10	25
`list_questionnaires`, `list_rfps`, `list_sources`	20	100

List results also carry the total count and the current page, so an assistant asked for “all of them” can page through rather than stopping at the first twenty.

When something goes wrong

Errors come back as a plain sentence the assistant can read out — a record that does not exist, a file that could not be fetched, a date in the wrong format, an action your role does not allow. Nothing fails silently, so if an assistant says it did something, the tool reported success.

An id belonging to another account reads as “not found”, because your connection cannot see outside its own account at all.