Skip to main content

August 2026

Model settings, versioning and sample import in AI Studio are now pages, and the newest version of a model can be rolled back. Calculation checks now handle negative amounts and totals of calculated fields.

Highlights​

AI Studio overhaul​

Model settings, version creation and sample import in AI Studio are now full pages that you can link to. The settings page has stacked sections, a Versions overview, and the model instructions next to the labels they apply to. The labels table supports inline editing, drag-to-reorder and search, and label order is saved. Creating a version now freezes the current state in place, and new work continues from it. Each version records where its samples came from, and the version page shows what will be frozen before you confirm. Sample import can draw from several data sources at once and shows the total first. The objectives list shows each model's architecture: retrieval, BERT, wordvector or conditional.

Rolling back a version bump​

You can now roll back the newest version of a model from the Versions overview on its settings page. The version becomes the editable working copy again. Samples added since the bump are carried back into it. Other changes to the working copy since the bump, such as edits to inherited samples, are discarded, and the confirmation dialog says so before you confirm. Connected projects keep pointing at the same version.

Availability

Only for retrieval models, only on the newest version, and only while nothing is connected to the working copy.

Week of 3 August​

AI & data​

Validation​

  • Negative amounts in calculations: amounts with a trailing minus, in brackets, or with a minus-like character from OCR are now read as negative. The 1.000,- shorthand still means a positive amount.
  • Unreadable figures in calculations: a figure that cannot be read now raises a warning instead of being left out of the sum, so some documents that passed before may now warn. Cells holding a dash or "n/a" are still skipped.
  • Totals of calculated fields: a total can now add up fields that are themselves calculated, such as line totals of quantity times price.
  • Size of a calculation error: a failed calculation now states the gap and its direction, for example Calculation is off by -3192.88. A minus means the field reads lower than the calculated value.

Fixes​

  • A document-level field is no longer counted twice when a calculation sums it with fields inside a group.
  • Correct decimal sums no longer fail on tiny rounding differences in addition, subtraction or multiplication.
  • A calculation margin written with a decimal comma, or entered as a negative number, now works correctly.
  • Import metadata filters no longer return nothing for some projects.
  • The organisation members page no longer crashes when a member has left the organisation.
  • Trained models no longer show pages that apply only to retrieval models.
  • AI Studio links elsewhere in the product open in the studio again and point to the right page.

Week of 17 August​

AI & data​

Export​

  • GZIP compression on HTTP connectors: an optional switch that compresses step export payloads, for receiving systems that reject large requests. It is off by default, and Test connector uses the setting too.

Fixes​

  • Opening an example from the diagnose dialog now loads the model version it belongs to.
  • An action refused because a version is locked now explains why, instead of reporting a permission problem.
  • Bumping or rolling back a version on models with older versions no longer fails.

Week of 31 August​

Validation​

  • Context in non-unique lookups: with Match unique row switched off, lookups now use the field's context fields too, so a value that validated before may now warn. A value repeated across many rows counts as one match.
  • Lookup metadata from matched columns: non-unique matches attach only the field's own column and its context columns, where all matched rows agree. To attach a full row, keep Match unique row on.
  • No-match message: a lookup driven by context fields that finds nothing now says no match was found in the reference data.

Fixes​

  • The description of Match unique row in the field and pipeline step dialogs now matches what it does.