Preparing Ask

Ask does not infer the business. It answers from what the data team wrote about each app, and that preparation is what separates a reliable answer from a confident wrong one. It all lives under Admin, on the Ask resources screen.

Only apps in Qlik's managed spaces appear for preparation. Content in personal or shared spaces stays out, because the portal works solely with what is published and governed.

Company context

A single Markdown text, applying to every enabled app. It is the place for the company description, the general business vocabulary, and the guidance that holds in any answer, such as currency and language.

It is also what supports questions asked without selecting an app. In that mode the agent knows the company context and the list of prepared apps, and decides which one answers.

The app context

Each app carries its own context document, divided into sections. The division is not organisational: each section reaches the agent with a different weight.

Section	What goes in	How the agent treats it
Context	What the app is and what it is for	Frames the answer
Business rules	Canonical measures, caveats, exclusions	Source of truth; the agent does not rediscover what is here
Assumptions	Fiscal year, currency, what counts as active, default exclusions	Applied by default and recorded in "How it was calculated"
Constraints and negatives	What to refuse, what to ask rather than guess	Inviolable: they prevail over the user's request
Vocabulary and synonyms	The user's term to the canonical measure or field	Resolves the term before exploring the data model
Examples	Typical questions, special cases	Reference for style

Business rules is the only mandatory section: without it the app cannot be enabled. The other five are what separates "answers" from "answers correctly".

Vocabulary is the most direct lever. The question arrives in the user's words, "revenue", "team", "portfolio", not under the measure's name. Without that map the agent resolves by closest name, and in an app holding Revenue, Net Revenue and Gross Revenue the right choice is not obvious. A few lines of vocabulary resolve more errors than pages of description.

The AI-drafted starting point

A button drafts a proposal for the rules from the app's name, its fields, its measures and the company context. It is a starting point for those who do not know where to start.

Until the text receives human editing the app stays marked as unreviewed, and users see, in that dashboard's chat, a notice that the answers may carry incorrect values. The notice clears once the text is reviewed.

Readiness assessment

The assessment reads the app in Qlik, meaning master measures, master dimensions, fields and sheet objects, together with the written context, and returns a list of points of attention graded by severity, along with a score.

What it commonly reports:

business rules missing, or too short to give direction;
assumptions and limits left undocumented, which is precisely what the agent gets silently wrong;
vocabulary left undocumented;
master measures without descriptions, or with similar names and no written rule on which to use for which question;
fields with hard-to-interpret names;
no validated queries, or queries that leave one of the three most common question shapes uncovered: a single figure, evolution over time, and a categorical breakdown.

The assessment does not block. An app holding critical points may go live, with the notice that answers tend to be worse until they are resolved.

Validated queries

A validated query is a ready-made slice of the app: dimensions, measures, filters and a row ceiling, under a name and a description. The agent prefers them when the question matches the description because they return the same formula every time, and appear by name in the "How it was calculated" panel instead of an expression assembled on the spot.

The query is built from the app's master items or from expressions in Qlik syntax, and is validated against the app before being saved: the portal runs it and displays the rows returned.

Two decisions deserve attention.

Master item or formula. Referencing the master item makes the query follow formula changes in Qlik. Copying the formula freezes the calculation, and a change made in Qlik never reaches it.

Starter question. A query may carry an example question, which becomes a clickable suggestion in that app's chat entry. It is the cheapest way to teach people how to ask.

Testing before publishing

An app that is enabled and not yet published is visible to administrators and to the users marked as testers. Publishing releases it to everyone, and unpublishing returns it to the previous state without erasing anything.

The preparation screen holds a test chat. It answers from the saved context, not from what is on screen, so save before testing again.

Privacy mode

Privacy mode has a default value for the workspace and a per-app setting.

With it on, no data row passes through the model: the query runs on the server and the server builds the chart. The model receives the result's structure, not its values. Users still request whatever format they want, and the answer's text runs shorter, because the model does not comment on figures it has not seen.

An app with its own privacy setting is queried only while in focus. It stays out of general questions, asked without selecting an app, because there the guarantee holds per conversation rather than per app.

Display

Four options govern what appears in each answer, across the whole workspace: the "How it was calculated" icon, the response time, the token total, and the loaded-context indicator. The default shows the first two.

Switching off the "How it was calculated" icon removes the user's audit of the number, and makes sense only where that audit happens by another route.

Usage and quota

The usage tab reports conversations per day, tokens, users, query calls and errors, broken down by app and by user. It answers two questions: which dashboards generate questions, and who is using them.

The question quota belongs to the workspace, not to the individual. The portal issues a notice as the period nears its end and blocks on exhaustion, with the request for an increase available from the notice itself.

Practices that pay

Write the vocabulary first. A few lines mapping the terms people use resolve more errors than pages of description.
Disambiguate similar measures. State which measure answers which kind of question. Where two look alike, the agent picks one with conviction, and it may be the wrong one.
Document the assumptions. Example: the fiscal year, the currency, what counts as active, and what is excluded by default.
State what the data does not cover. Periods, units, granularity. The agent then stops answering beyond what it knows.
Cover the three question shapes with validated queries. A single figure, evolution over time, and a categorical breakdown.
Review the draft generated by the AI before publishing. It gets the structure right, but the business rules and vocabulary need human eyes. Reviewing the draft is highly recommended.
Test with the questions people actually ask, not with the ones the app answers well.

The limits of preparation

Context directs the agent; it does not repair the data model. A wrong measure in the app stays wrong in the answer, and the correction happens in Qlik.

Preparation also grants no access. Every query runs under the permissions of the person asking, and an app published to Ask still answers only what that person would see by opening the app.