Version 8.4.0

Minor release: statistics data in, forecasts and data-driven dashboards out. A class diagram can be imported from any SDMX statistics API, an NN diagram generates a runnable time-series forecaster, and charts and metric cards can read from any backend endpoint. GUIs designed by the modeling assistant now keep their styling and inline SVG in the generated React app. No breaking metamodel or API changes: DataBinding.endpoint is optional, and models without it generate the same pages as before.

Importing data

  • A class diagram from an SDMX dataflow. The new besser.utilities.sdmx_to_buml utility reads an SDMX 2.1 dataflow’s structure, from a provider’s API or an XML file, and builds a validated B-UML class diagram. Dimensions become enumerations or classes, code hierarchies (HS2 to CN8, regions) a parent/children self-association, and each class keeps its SDMX role in its metadata. 14 providers are preset, among them Eurostat, STATEC LUSTAT, the ECB, the OECD, the IMF and the ILO. See From an SDMX dataflow to B-UML.

  • One “From Data” import in the editor. The Project Hub’s “From spreadsheet” step becomes From Data, with Files (CSV, XLSX) as before and Statistics API (SDMX): pick a provider and a dataflow id. A failed import no longer leaves an empty project behind, for CSV too.

  • Safe fetching on a public server. The backend endpoints POST /sdmx-to-domain-model and GET /sdmx-providers fetch over https to public addresses only, checked again on every redirect and pinned to the checked IP. Each import has one deadline and a byte and request budget, XML declarations are refused, and at most 2 imports run at a time. Custom URLs are off unless BESSER_SDMX_ALLOW_CUSTOM_URL is set.

Generators

  • Time-Series Forecaster. TimeSeriesForecasterGenerator turns a B-UML NN model and a small ForecastConfig (window, horizon, season, log, differencing, calendar features, standardization, seeds) into one runnable PyTorch module. The module handles windowing and its inverse transforms, seeded training on one series or several, predict, save/load, a rolling-origin backtest against the seasonal naive, and a CSV command line. Every layer the PyTorch generator supports can be used. In the editor: Generate → PyTorch → Time-Series Forecaster on an NN diagram. See Time-Series Forecaster Generator.

  • Charts and metric cards that read from an endpoint. The new DataBinding.endpoint lets a chart series, a chart’s own binding or a metric card read rows from a backend path (/stats/sales) or an http(s):// URL instead of its class’s CRUD list, so a chart on a large table no longer downloads every row. It goes through the metamodel, the .py export, both editor converters and React and web app generation. In the GUI editor, the series and the metric-card settings have an Endpoint field.

  • Charts built in Python keep their series in the editor. The B-UML to JSON conversion never exported chart series, so a chart built in Python lost its series and bindings when opened in the editor.

  • Assistant-designed GUIs keep their styling in the generated app. The modeling assistant writes CSS classes as HTML class attributes, which the GUI conversion did not read, so generated React apps had no className and came out unstyled. Inline SVG (icons, illustrations) was generated as <div> elements and drew nothing. Both are fixed: classes are read from either place and exported once, and SVG elements keep their tags with React attribute names.

  • No HTML injection through GUI attributes. The React generator no longer passes dangerouslySetInnerHTML, ref or key from a GUI component’s attributes into the generated page.

B-UML export

  • Attribute multiplicity and read-only in .py exports. Exported B-UML scripts did not write either, so emails: str [0..*] came back as a single value and a read-only attribute as writable, in the editor’s .py import as in plain Python. Generators that read attribute multiplicity (Spring, JSON Schema, SQL) then produced different code from the reloaded model. Both are now written; attributes with the defaults are exported as before.

Deploy alongside

Deploy the modeling agent with this release:

  • A GUI from a screenshot or mockup (BESSER-PEARL/modeling-agent#38, #643). An image attached in the assistant chat is read as a page mockup when the message asks for a GUI, or, with no message, when it shows a user interface. The assistant builds a GUI with the same sections, text, colours and fonts, with tables bound to classes when a class diagram exists. Several screenshots become one app with a page per image. Diagram images are still read as diagrams, and the message can now name the diagram type. Reading images needs a vision model: the server’s OpenAI models or the user’s own OpenAI or Anthropic key.

  • Inline SVG in assistant-designed GUIs shows in the editor canvas instead of as empty boxes.

  • A table, chart or metric card no longer shows its title twice: once as the section heading and once as the widget’s own title.