Paginated Reports (RDL)

A paginated report is a print-faithful, multi-page document (invoice, statement, operational list, regulatory filing) defined by a single .rdl XML file. The .rdl is plain, hand-editable, diff-friendly XML holding everything: data sources, datasets, parameters, page setup, layout, and expressions. Power BI Report Builder is a Windows GUI over this same XML, so a coding agent edits the artifact directly.

This skill teaches the RDL format and its unwritten rules, how to connect to data, the rendering quirks, and a dev loop that confirms a report actually renders. It is the home for everything RDL. For interactive screen-first reports (PBIR), use reports:create-pbi-report instead; see "Is paginated the right tool?" in references/report-structure.md. This is for Power BI paginated reports (the Power BI service / Fabric), not classic SSRS; the format is shared but the environment differs (see references/differences-with-ssrs.md).

Before building: interview the user

A paginated report is a fixed-format document where the data source, parameters, and delivery target are expensive to change later. Before authoring anything, run a requirements interview using references/questionnaire.md. Treat it as a dynamic, two-way conversation, not a form: research the data source yourself (inspect it, run candidate DAX, read a similar example), bring concrete options and a quick Enter Data draft for the user to react to, and follow the threads that matter rather than reading a fixed list. Settle the essentials (purpose, data source, target workspace and capacity), reflect back a short brief, then build against it and refine the rest as the draft takes shape.

Three rules that prevent most breakage

1. Element order is load-bearing in practice. Report Builder and the report processor's reader expect the conventional order and fail to load out-of-order children with no useful error (the failure is from the processor, not schema validation, so an XSD validator will not reliably catch a reorder). Preserve the documented order when editing. This is the number-one cause of a broken hand-authored .rdl.
2. Validate after every structural edit. Run scripts/validate_rdl.py <file.rdl> to catch order, name-collision, tablix-count, reference, and unit errors before publishing. A clean pass plus a successful render is the bar for "done".
3. Reuse a template, do not hand-type from scratch. The verbose XML (charts especially) is error-prone to write by hand. Copy the closest assets/*.rdl starter and change the data source, query, fields, and layout. Regenerate rd:ReportID to a fresh GUID when copying.

Workflow

1. Pick a starting template       assets/enter-data-starter.rdl | semantic-model-starter.rdl | sql-starter.rdl
2. Iterate layout offline          edit XML; keep an Enter Data dataset so no live source is hit
3. Validate                        python3 scripts/validate_rdl.py report.rdl
4. Wire the real data source       swap <DataSource>/<DataSet><Query>; keep field names so layout is untouched
5. Validate again                  python3 scripts/validate_rdl.py report.rdl
6. Publish to a workspace          scripts/publish_rdl.sh report.rdl <workspaceId>
7. RENDER AND CONFIRM (hard gate)  scripts/export_rdl.sh <reportId> <workspaceId>  ->  open the PDF

Steps 2-3 are local and free. Do the layout work against the Enter Data inline dataset (a fixed handful of typed rows embedded in the .rdl) before touching a live source; re-querying a real source on every layout tweak is the biggest time sink. Keep the Enter Data field names identical to the real query's fields so swapping the source touches only the <DataSource> and <DataSet><Query>, never the layout. Publishing and rendering need a workspace on Premium/Embedded/Fabric capacity.

Step 7 is the real bar, not step 5. A green validate_rdl.py is necessary but not sufficient: it checks structure, never expressions, field references, or DAX, so a report with a wrong DataField, a mistyped measure, a @-prefix mismatch, or a parameter pointing at the wrong column passes validation, publishes, and only fails (or renders blank/wrong) at export. So the moment the real source is wired (step 4), render immediately and: confirm the PDF opens with the expected, non-empty content; read the export error body if it fails; and verify the data contract by running the exact EVALUATE/SQL against the live source (via semantic-models:dax) and checking the returned column names and types match the <Field> definitions. Do not call a report done on a green validator alone.

Starting templates (assets/)

• enter-data-starter.rdl: portrait-letter report with a title header, page-number footer, and a 3-column table bound to an embedded Enter Data dataset. No data source needed; renders in the service with zero config. Best starting point for layout iteration and the fastest thing to test end-to-end.
• semantic-model-starter.rdl: connects to a Power BI semantic model (PBIDATASET) via DAX, with a single-value Category parameter wired through TREATAS (plain valid DAX; render-proven against a live model) and a dataset-driven default. Fill the REPLACE_WITH_* tokens (dataset GUID, workspace/model names, table/column/measure names).
• sql-starter.rdl: connects to Azure SQL (SQLAZURE) via T-SQL with a multi-value Category parameter. Fill the server/database tokens and adjust the query/fields.
• platform-template.json: the .platform sidecar Fabric Git expects beside an .rdl in a <Name>.PaginatedReport/ folder.

Scripts (scripts/)

• validate_rdl.py: stdlib-only structural validator (cross-platform). Checks XML well-formedness, the 2016 root namespace, a valid rd:ReportID GUID, top-level element order, Name uniqueness, tablix column/row/cell-span invariants, dataset-to-datasource and tablix-to-dataset references, embedded-image references, and dimension unit suffixes. It does not check expressions or live field references; those surface at render time. Run after every structural edit.
• publish_rdl.sh / publish_rdl.ps1: upload an .rdl via the Power BI Imports API (a multipart post fab api cannot do) and poll the import to completion. When the conflict mode is omitted they auto-detect Abort (new report) vs Overwrite (existing), and surface the API error body on failure. Token comes from $PBI_TOKEN or is minted inline via az; never written to disk. Usage: publish_rdl.sh <file.rdl> <workspaceId> [Overwrite|Abort] [displayName].
• export_rdl.sh / export_rdl.ps1: the render-to-verify step. Trigger an export-to-file (PDF/XLSX/etc.), poll, and download the result. Same token handling. Usage: export_rdl.sh <reportId> <workspaceId> [format] [outfile]; set PBI_PARAMS to a JSON array to pass report parameters. (Status endpoint is GET exports/{id}, not .../status.) The .ps1 variants are for Windows; validate_rdl.py is Python and runs everywhere.

Connecting to data

A paginated report embeds its own data sources and datasets (the service does not support shared .rds/.rsd). Each <DataSource> sets a DataProvider; each <DataSet> is a query plus the fields it returns.

Parameters wire across three places that must agree: the <ReportParameter>, the dataset <QueryParameter Name="@x"> whose <Value> is =Parameters!x.Value, and the query's @x reference. Give every parameter a default so the report opens without forcing a selection; back dropdowns with a values dataset; keep cascades shallow. Author the DAX with the semantic-models:dax skill (or reports:pbir-cli's model -q) against the live model so it is verified before it goes in <CommandText> — do not hand-roll it. For a semantic model, a single-value parameter via TREATAS is the robust, render-proven pattern (it is what the starter uses); multi-select needs RSCustomDaxFilter, which is a fragile Report-Builder-generated construct (see the caveat in references/data-sources.md). Full connect-string anatomy (including the non-negotiable sobe_wowvirtualserver-<guid> and Integrated Security=ClaimsToken for PBIDATASET) and the Enter Data block are in references/data-sources.md.

Rendering quirks to anticipate

• Preview never equals PDF. Report Builder and the service browser always preview through the HTML (soft page-break) renderer. Page counts, margins, and header height differ from a PDF/Image export. Verify the actual delivery format by exporting to it, not by trusting preview.
• The PDF width trap. Body <Width> + left + right margins must be <= PageWidth, or hard renderers emit blank pages to the right. Landscape Letter with 0.5in margins gives 10in usable width.
• Two execution environments. In the service a report runs "standard" or "optimized" automatically; the assignment depends on whether every RDL expression is in the optimized subset. The cost of falling to standard lands at startup (container reuse), not the per-row loop, so regaining the optimized path is not a general speed fix; profile before chasing it. VB date/string transforms inline in the RDL (e.g. Weekday(...)) force standard; move such computation into the query/model. View > Diagnostics names the exact non-optimized expressions.
• Scope-bound expressions. Globals!PageNumber/TotalPages work only in page headers/footers. RDL dataset/tablix filters apply after the full fetch, so they do not reduce data pulled; filter in the query.

Full renderer taxonomy, the standard-vs-optimized detail, pagination control (page breaks, KeepTogether, KeepWithGroup, RepeatOnNewPage), and expression scope rules are in references/renderers.md.

Structuring a report

Lock the physical page frame (size, orientation, margins, usable width) before placing anything; every column width and font size is constrained by it. Build from a small set of regions: page header (title, logo, run date, parameter echo), body (the data region), page footer (page numbers, confidentiality). Choose the data region by data shape: table for fixed columns, matrix for data-driven columns, list for repeating free-form blocks, all of which are the same <Tablix>. Make pagination deliberate: group by the natural document unit, break between groups, and repeat headers across pages. Design guidance specific to paginated reports (including working from a mockup) is in references/report-structure.md.

Reference files

• references/questionnaire.md: the intake interview to run before building (source, audience, purpose, parameters, distribution, design, workspace/capacity, Report Builder access) plus sensible defaults.
• references/differences-with-ssrs.md: how Power BI paginated reports differ from classic SSRS (same RDL, different environment); what to convert in SSRS-era samples.
• references/rdl-format.md: the standalone RDL XML format reference: namespaces, top-level order, data sources, datasets, ReportSections/page/header/footer, the full Tablix structure (flat and grouped), textboxes, images, charts, parameters, and the unwritten rules and common breakage. Read this before hand-editing RDL.
• references/data-sources.md: connecting to semantic models, SQL, Analysis Services, and other sources; connect strings, query languages, field naming, parameter binding, and the Enter Data connector.
• references/renderers.md: rendering extensions, standard vs optimized execution environments, pagination control, and expression scope quirks.
• references/report-structure.md: designing a good paginated report: page frame, region skeleton, choosing the data region, deliberate pagination, parameter UX, working from a mockup.
• references/dev-loop.md: the author/validate/test/publish loop, Enter Data iteration, PBIRS local rendering, publishing, and the Fabric Git on-disk layout.
• references/example-reports.md: curated real-world .rdl files on GitHub (Microsoft samples, a PBIDATASET example, parameters/matrix/sparkline examples) to copy verbose constructs from rather than hand-typing.

Related skills

• semantic-models:dax: author and validate every DAX query a PBIDATASET dataset will run, against the live model, before pasting the EVALUATE into <CommandText>. Do not hand-roll DAX for a paginated report.
• fabric-cli: references/paginated-reports.md there documents the full Power BI REST surface for paginated reports (datasource management, gateway binding, RLS, admin, and the export-to-file options). Use it for everything past publish that the bundled scripts do not cover.
• reports:pbir-cli: an alternative way to discover model fields and run DAX (model -q) when authoring queries.
• reports:create-pbi-report: when the user actually wants an interactive report, not a paginated one.