Skip to main content

July 2026

Extraction validation now runs automatically, duplicate uploads can be stopped before they create tasks, and grouping models can be trained from the interface.

Highlights​

Live extraction validation​

Validation now runs automatically about half a second after you edit an annotation, so the results on screen always match the document. The Validate changes button is gone. External lookup responses are cached, so validating as you type does not add load on your systems. Manual values and lookup selections are kept as you edit, and when a lookup returns several similar matches, the one you pick stays selected.

Duplicate filtering​

Duplicate filtering is an opt-in setting in a project's import settings. You set a time window (1 to 99 seconds, minutes, hours or days) and, optionally, a folder for duplicates. An upload whose files all match a recent upload by name, type and size is flagged as a potential duplicate and no tasks are created. The upload stays visible, with a warning in the activity panel and a notification event.

Grouping model training from the interface​

Lock a grouping version and request training to start a training run; the interface shows its progress. Grouping rules can now find where one document ends and the next begins inside a single file, such as a scanned batch of documents of different lengths. Each run reports a score measured on samples held out from training.

Week of 6 July​

Validation​

  • External lookup translations: values returned by a lookup API can now be translated.
  • Lookup endpoints per environment: a test environment can now call a different lookup API URL from production.
  • Row editing in lookup lists: you can now edit rows in a lookup list.

Fixes​

  • Flow grouping in single-step projects works again.
  • The autocomplete action is no longer active when autocomplete is switched off.

Week of 13 July​

AI & data​

Prediction & processing​

  • Edits clear the training mark: changing a label on an approved task now clears "Marked for training".

Interface​

  • Activity panel: long activity branches collapse, so tasks that need action are easier to find.

Fixes​

  • Un-approving a task no longer removes its annotation from a locked sample set.

Week of 20 July​

Validation​

  • Live extraction validation: described above.
  • Up to three description columns: the autocomplete subtitle of a lookup list can now combine up to three columns.
  • Validation tab for classification: classification tasks now have a Validation tab listing the project's classes with a confidence for each, and who or what assigned the selected class.

Import & document processing​

  • Forwarding keeps original files: forwarding a task to another project now carries the original email with it.
  • When every attachment is filtered: such an upload can now either stay where it is with a warning, or follow the filtering rules.
  • Filtered files summary: filtered files collapse into one summary on the upload card, with undo per file.

Interface​

  • Leaderless groups: in the template editor you can insert a group first and set its leading entity later from the group header.

Fixes​

  • Business Rules that read numbers from currency or text fields, or that return 0, now apply.
  • Generated Business Rules no longer fail on documents with empty numeric fields.
  • Forwarding a task no longer creates duplicate tasks in the receiving project.
  • A task no longer shows as exporting while other tasks in its upload are still in review.

Week of 27 July​

AI & data​

  • Provider failover queue: choose which AI providers a model may use and in what order. Anthropic is the primary provider and cannot be switched off.
  • Model instructions as a list: instructions are now an ordered list of lines you can switch on and off one at a time.
  • Mark samples for review: flag samples for review one at a time or in bulk, and filter the sample list by the flag.
  • Inline tag management: create and edit tags directly in the tags table.