Tableau content search strategy
Use the stages that the active Tableau MCP server supports. Tool names, parameters, permissions, and returned fields may vary.
1. Resolve project scope
When the user names a project, search or list projects to resolve the canonical project identity and hierarchy. Use descriptions and parent metadata as evidence. For a fuzzy or owner-based hint, present a small candidate set instead of silently selecting a project.
When the user asks to browse projects, prefer top-level results first and drill into children on request. Do not assume that wildcard, owner, or parent filters exist; inspect the active tool schema and post-filter returned metadata when necessary.
2. Choose search breadth
- Small scoped project: a bounded inventory of views/workbooks plus relevant datasources can maximize recall.
- Large scoped project: lead with server-side content search and supplement only when results are sparse or naming is opaque.
- Full site: use bounded server-side search with a few deliberate term sets. Do not download a complete site inventory by default.
If project size is unknown, start with a bounded search rather than making a separate expensive counting pass.
3. Search concepts
Extract:
- primary nouns and business concepts;
- implied measures and dimensions;
- time intent such as trend, monthly, YTD, or historical;
- conservative synonyms.
Useful starter relationships include:
| Concept | Related terms |
|---|---|
| sales | revenue, orders, bookings, ARR, MRR |
| customers | clients, accounts, users, subscribers |
| churn | attrition, retention, cancellation |
| profit | margin, earnings, net income, EBIT |
| pipeline | funnel, opportunities, deals, prospects |
| performance | KPI, metrics, scorecard, health |
| trend | time series, historical, monthly, weekly |
| region | geography, territory, market, location |
| product | SKU, item, catalog, offering, category |
| employee | headcount, staff, workforce, HR |
Do not treat every related term as equivalent. Add domain-specific synonyms supplied by the user or established in the current conversation.
4. Layer evidence
Start with content names, descriptions, tags, project context, and types. Add field metadata only for promising datasources or the datasources behind leading views. Add view/workbook detail when it provides a canonical URL, usage metadata, upstream datasource evidence, or companion views.
Pulse definitions are in scope only when the user asks about a metric/KPI or requests a broad sweep and the server exposes Pulse discovery.
5. Handle degraded access
- Search unavailable: use bounded list operations if available.
- Datasource listing denied: use upstream datasource references returned with views.
- Field metadata unavailable: retain name/description evidence and lower the certainty of field-level claims.
- Pulse unavailable: omit it from searched counts and name it as unavailable only when it was requested or material.
- Canonical URL unavailable: omit the link instead of constructing one.
Separate an empty successful search from a permission or availability failure.
6. Evidence normalization
Build candidate records only from returned metadata. Deduplicate identical canonical IDs before ranking. If the same object appears through several searches, merge non-conflicting evidence and retain the server-returned canonical link.
Use an explicit as_of date in the ranking input when applying the 90-day recency signal. Without it, the ranker omits recency so repeated runs remain deterministic.