Repository navigation
User Guide
For people using a Tableau dashboard with the OpenVizPilot chat. German version: Benutzerhandbuch. Back to Home.
The chat answers questions about the currently open dashboard: metrics, filters, parameters, your selection, data sources, and where a number comes from. It only reads data you're already allowed to see in Tableau — row-level security and user filters apply unchanged, because every query runs in your own Tableau session.
What it does not do:
- change the dashboard on its own — filters, parameters or highlights are shown at most as a suggestion (see section 7) and only applied once you click it.
- answer questions with no connection to the dashboard — an off-topic question gets a fixed refusal: "This question is outside the dashboard context. I only answer questions about the open dashboard and its data."
- show data your Tableau permissions don't allow.
The chat panel is part of the dashboard — the author placed it there, so you don't install anything yourself. To use it, an administrator has to grant your account AI chat access (and, for some dashboards, Tableau API access as well, when server-side features are used); you can see the current status in the panel under "Approval status".
- Open the dashboard in Tableau as usual.
- Pick the mode above the input field: Ask for individual questions and answers, Investigate for a multi-step root-cause analysis (section 4).
- Type your question and send it — or start from one of the suggested prompts, e.g. "Summarize the dashboard briefly." or "Which filters are active right now?".
With an Enterprise license, the chat asks on your first visit to a dashboard which answer focus you'd like (for example a management summary, a detailed analysis, compact tables, recommendations — or none); without the license, the starter suggestions appear directly.
The more specific the question, the more targeted the answer. Examples that actually work:
- "Summarize the dashboard briefly." — reads the key metrics and active filters and answers with a few core statements plus a compact table.
- "What stands out in the data? Name the three most important points." — aggregates the full summary data (not just the visible first page) and names outliers with figures, not impressions.
- "Which filters are active right now?" — lists field, filter type, selected values/range and exclude mode per worksheet.
- "What have I selected right now?" — reads your selected marks; if nothing is selected but something is highlighted (highlighter, legend click, highlight action), it uses that and says so.
- "Where do I change the region?" — names the actual control on the dashboard instead of guessing a field.
- "What does 'Order details' show?" — if that worksheet's zone is currently hidden behind a show/hide control, the chat says so instead of describing something you can't see, and offers to show it.
- "Where does this number come from?" or
/lineage Revenue— returns the full provenance chain (section 6).
Above the input field you can switch between two modes:
- Ask — for individual questions with a direct answer.
- Investigate — for "why" questions like "Why did margin drop in the South?". The chat first writes a 3–6 step analysis plan, works through it with more tool calls (breakdown by dimension, period-over-period comparison, aggregations), and closes with Main cause, Evidence and Sources — every claim backed by a visible tool result.
With an Enterprise license and your personal "server data" approval, Investigate mode also shows a scope switch: Only this dashboard or Entire Tableau environment. In the second case the chat also searches other workbooks/views, reads at most five of them server-side on your behalf, and still closes with Main cause, Evidence and Sources (per source: workbook · view · link). Without the license or your personal approval, the chat explains that plainly and investigates only the open dashboard instead.
Type / in the input box to open the command menu. Every command exists in German and English (except /top, which has the same name in both), and an administrator can replace these defaults or add their own per dashboard.
| Command (EN / DE) | What it does |
|---|---|
/summary · /zusammenfassung
|
Management summary of the dashboard: 3–5 key statements plus a metrics table. |
/anomalies · /auffaelligkeiten
|
Finds the three biggest anomalies (outliers, unusual ratios) with drilldown. |
/compare <A> <B> · /vergleich <A> <B>
|
Compares two segments, regions or time ranges in a table (metric · A · B · difference absolute/%). |
/top <N> <Dimension> |
Top-N ranking with each entry's share of the total and a note on concentration. |
/actions · /massnahmen
|
Up to three prioritized recommendations, each backed by evidence (figure + source). |
/report · /bericht
|
Formatted short report ready to copy: heading, overview, metrics table, findings, recommendation. |
/data-quality · /datenqualitaet
|
Checks for obvious data gaps and whether active filters/parameters hide part of the data. |
/exec-brief <audience> · /fuehrungsbericht <Zielgruppe>
|
Executive brief with a fixed structure for a given audience, including the data basis (worksheets, filters, as-of state). |
/lineage <metric/field> · /herkunft <Kennzahl/Feld>
|
Provenance chain of a number: dashboard → worksheet → field → catalogue definition → data source → upstream tables → active filters → certification status. What an API doesn't provide is marked "not available", never invented. |
Every answer keeps its analysis trail: an expandable section shows which tools ran on which worksheet, down to the raw tool result — so a figure can be traced instead of believed.
Two kinds of domain knowledge feed into answers as well:
- Glossary: the dashboard author can maintain a dashboard context/glossary (KPI definitions, abbreviations, domain notes), visible to everyone using that workbook. Ask "What does contribution margin II mean?" and the chat uses this house definition instead of textbook knowledge.
-
Metric catalogue (if an administrator maintains one under
/admin→ "Kennzahlen": definition, synonyms, owner, data source, verified questions): if you ask about a metric listed there, the chat looks it up first. The answer then carries the "Verified definition" badge — shown only when the lookup actually happened, never as a bare claim.
With the "actions" license feature, a question like "Show me only the North region" gets a proposed dashboard change as a chip — apply or clear a filter, change a parameter, highlight marks (e.g. "show me the top 3 regions"), or show/hide a dashboard zone. Important points:
- The chip always spells out the technical change in plain text.
- Nothing runs on its own — only your click on the chip applies the change. This is deliberate (human-in-the-loop) and also reduces prompt-injection risk.
- Chips set categorical filter values; date ranges are still chosen directly in the dashboard.
- Without this license the model isn't told the action syntax at all, so it never proposes one.
Both features need a personal user identity from Tableau, and therefore Tableau 2023.2 or newer; on older versions the extension still runs, but the settings panel explains plainly why the personal sections are missing.
- User memory: the chat can remember personal details (name, role, preferred views/formats) to personalize answers for you. A lightweight model extracts these facts exclusively from your own messages — dashboard data never reaches this extraction, and the prompt explicitly forbids storing it. In the settings panel under "Saved info about me" you can see every stored fact and delete it anytime — this right applies regardless of whether the license is currently active; facts saved during a time with a license stay deletable even after no new ones are added.
- Saved queries: for each dashboard you can pick an answer focus (e.g. management summary, detailed analysis, compact tables, recommendations) and save up to five of your own starter questions as chips (star icon on an answer).
With the "watch" license feature you can ask the chat to check a metric on a schedule, e.g.: "Let me know when margin drops below 20%." The chat never creates a rule on its own — it shows a confirmation card with every field editable (name, threshold, schedule, channel); only your click on "create rule" creates it.
Prerequisites, re-checked on every run: the license (which requires "server data", the Tableau Server integration and SSO), your personal "server data" approval, your consent to server-side data access, and the site-wide "allow server-side data" switch. If any of these fails, the rule is automatically disabled with a plain-text reason.
- Schedule: every 15 minutes, hourly, every 6 or 24 hours, or weekly (weekday + hour, with a time zone).
- Condition: value below/above a threshold, or a percentage change from the previous run.
- Notification: sent only on a state change — an alert when it crosses into the flagged state (and again if it stays flagged uninterrupted for 24 hours), an all-clear when it returns to normal. Delivered via webhook, Microsoft Teams, or email (if enabled by the administrator), each with a link to the view.
- Managing rules: the settings panel lists your own rules with an on/off toggle, "test now" and delete. There's no full rule-creation form — new rules are only created via a chat proposal.
- The rows that were read are never stored, only the last computed value.
- What goes to the language model: your question, plus the dashboard data that model needs to answer it — limited to what you can already see in Tableau. Row-level security and user filters apply unchanged, because every query runs in your own Tableau session.
- The middleware (the server behind the chat) has no Tableau identity of its own, caches no dashboard data, and keeps no chat history — your conversation lives only in your browser. What it logs server-side is metadata only, never the content of your questions or dashboard data.
- What is stored server-side is only the application's own state: accounts/sessions, admin settings, anonymous usage counters, and — with an Enterprise license — your personal memory facts and saved queries.
- Usage counters are non-reversible pseudonyms — never names, IDs or question content.
- Server-side access to other views (Watch, estate-wide investigation) needs explicit opt-in: a site-level switch, a personal approval, and your consent. Without that, everything stays in the browser as described above.
- Your saved memory facts can be viewed and deleted at any time in the settings panel.
- "Not granted" / approval pending: your account needs "AI chat" approval (and, depending on the dashboard, "Tableau API" approval too) from an administrator, plus a personal sign-in via SSO or a user account. Contact your administrator.
- An Enterprise feature is missing (actions, memory, saved queries, Watch, estate-wide investigation): these features require a valid Enterprise license. Without one, the chat keeps working normally, just without these extras — ask your administration about the license status.
- "This question is outside the dashboard context": that's by design, not a bug — the chat only answers questions about the open dashboard. Rephrase your question with a clear link to the dashboard.
- "Dashboard changed": a note in the header means the worksheets or layout changed since the last answer — just ask your question again so the chat answers with the current state.
- Personal sections (memory, saved questions) are missing: these need a user identity that's only available from Tableau 2023.2 (Extensions API 1.11) onward. On older versions, the rest of the chat still works.
-
Backend URL is rejected: only HTTPS addresses (or
http://localhostfor development) are allowed. - General help: the "Help" link in the settings panel opens this wiki.