Ask about the dashboards

The search box at the top of the gallery takes a sentence, not keywords. Ask “anything on MAC table usage?” or “which dashboards suit a help desk technician?” and you get a written answer plus the dashboards it is about, as cards you can open.

It answers only from this gallery. The model behind it is shown nothing about these dashboards except the rows a search returned, so it has no general knowledge of them to fall back on — and when nothing fits, it says so rather than offering the nearest thing as if it fit.

What it can answer

• “Is there one for X?” — the common case, and the one the retrieval is tuned for.
• “What is here for a network engineer?” — a question about a category. The reply names a few and states the real total, and the cards carry the whole set. See Category questions return the whole set.
• “Anything on err-disabled ports?” — a question about a metric, which is usually the name of a chart rather than of a dashboard. Chart titles and section headings are read out of each export bundle and indexed, so these land.
• “How many are tagged Wi-Fi?” — a question about a size. The answer engine is given the real total, not the number of rows it happened to retrieve.

Questions are capped at 800 characters.

How a question is answered

Reading that as sentences, because each relationship matters on its own:

1. Your question goes to a search tool, never straight to an answer. The model turns it into keywords and an optional category, and calls search_dashboards.
2. Two rankings run over the same candidates. One is a full-text ranking whose columns are weighted — a dashboard carrying fifteen chart titles must not outrank one whose own title is an exact match. The other compares your question against a stored vector for each dashboard.
3. Your own sentence is embedded as well as the model’s rewrite, and the two semantic rankings are fused with each other before they meet the full-text half. The words a rewrite drops are often the ones that would have found the row.
4. Reciprocal rank fusion combines the two halves. A dashboard has to place well in at least one of them to survive, which is where precision comes from.
5. Only published dashboards are reachable. The published check is written into the search’s own SQL, again into the query that loads the vectors, and again when the rows are finally read — so a candidate cannot arrive through the vector table and skip it. There is no argument the model can write that reaches an unpublished draft.
6. The written answer and the cards come from the same rows. The cards are selected from what the search returned; they are never parsed out of the reply.

Why the cards cannot be wrong

The prompt asks the model not to invent a dashboard. The card selection is what makes it not matter if it does — a dashboard that does not exist has no row, so it cannot get a card. Anything you can click on is something the search actually found.

Every question is a fresh one

The endpoint accepts no conversation history. It is not validated and rejected — it is simply never read. Anyone can post any body to a public endpoint, and a forged previous answer is indistinguishable from one this site really produced, which would make every rule about what may be asserted negotiable.

So the box takes one question and answers it, and the interface matches: one answer on screen, replaced by the next. Asking a follow-up means asking a whole question — “and which of those is for switching?” has nothing to refer back to.

Category questions return the whole set

Asking what a category contains is a question about a set, so a category search returns the entire tagged set rather than the first few. Two things follow:

• The reply states the real total. It names a few of them in prose and tells you how many there are. Before the total was available, the answer engine reported the number of rows it had retrieved — five — as though five were the answer, when sixteen matched.
• Every one of them still gets a card, even where the prose names three. The cards are how you reach the rest of the set.

When an answer’s cards are only part of a category, a chip appears under them:

Show all 16 in the gallery →

It appears only when there is a rest to show — if the cards already are the whole category, there is nothing to offer. The point of it is the URL: an answer has no address, and a filtered gallery is a link you can send.

What you see while it works

A question takes a few seconds, and the panel reports what is happening rather than showing a bare spinner: which search is running, the terms it is running, and how much matched. The answer itself arrives once, at the end.

Limits worth knowing

Limit	Value
Question length	800 characters
Searches per question	3
Model turns per question	4
Wall clock for one model call	40 seconds
Answer length	1,200 tokens
Questions per minute, per address	20

Two different models answer, depending on how much of the day’s budget is left — they were measured as equal on recall, so the swap costs precision rather than coverage. Once the day’s allowance is spent the box stops answering and says so, and everything else on the site keeps working. Both are covered in Limits and quotas.

When search is resting, the filters are not

The category rail and the dashboards themselves have nothing to do with the model. If the search box is unavailable, filtering still finds the same dashboards — see Browse and filter.

Next steps

Request a dashboard When the honest answer is that nothing here fits, the answer panel offers to pass the request on.

Limits and quotas The daily ceiling on model spend, what happens past it, and every rate limit.

Troubleshooting The exact sentences the search returns when it refuses, and what each one means.