Email and hotfolder import, views for the transaction list, and review thresholds

Documents arrive on their own

Two new connectors bring documents in without anyone uploading them.

Import from email watches a mailbox and turns each message into one transaction with its attachments. Microsoft 365 inboxes connect through Microsoft Graph, and any other provider connects over IMAP. See import from email.

Import from a hotfolder watches a folder on a self-hosted deployment and picks up files as they land, which is the simplest way to connect a scanner or an existing file drop. See import from hotfolder.

You can also monitor how your connectors are running in production, along with how many samples have been collected for few-shot learning.

A demo project that routes email by intent

There is a new demo project in the shared library that reads incoming email, works out what the sender actually wants, and routes the message accordingly. Open the library and look under Marketplace to copy it into your own tenant.

The video at the top of this page walks through the whole thing end to end.

Views for the transaction list

The transaction list now has views. A view defines which columns are shown, how the list is filtered, and how it is sorted, both while you design a project and while people work in it. Anyone can build and save a view for themselves, and a project designer can publish views for the whole project or for specific roles. See views and columns.

Two related improvements:

• A column can now combine several fields, for example a document number that lives under a different field name in each document type, or the value of whichever metadata key is present.
• If you name your transactions in the workflow settings, for example “loan application” or “email”, the new transaction button uses that name, so the screen speaks your users’ language.

Decide what needs a human

The review step now takes your own confidence thresholds for splitting, classification, and extraction. Anything below the threshold you set is flagged for review and highlighted in yellow on the review screen, so reviewers spend their time where the system is genuinely unsure. See the review step.

The transaction table gained matching columns: issues left to review, issues before review (a fixed record of what the run started with), and issues corrected.

A clearer export schema

The JSON and XML exports have been revised. The new schema is much more straightforward, and it can export any state of a transaction, not just the final one. Exporting both the final and the pre-review state lets you build your own analytics on exactly what review changed.

The change is backwards compatible. Existing projects keep the previous schema and keep working, and you can move a project to the new version when you are ready. New projects start on the new schema. See the JSON output options.

Monitoring, reporting, and the work queue span projects

Monitoring, reporting, and My Work Queue are no longer limited to one project at a time. You can watch several projects on one screen and build reports across them, and reviewers can pick up or be assigned review tasks from more than one project. See monitoring and reporting.

Digitally born PDFs skip OCR

When a PDF already contains the exact text a reader sees, the system now reuses that text instead of sending the page through OCR. It is faster, cheaper, and more accurate on generated documents such as invoices and statements, and the activity log records where each page’s text came from. See the PDF text layer setting.

System administration for self-hosted deployments

Self-hosted deployments now have a published CLI and matching skills for coding agents such as Claude Code. Use them to set the application URL, turn public sign-up off (you usually do not want it on a self-hosted install), create and delete tenants, and more, without a redeploy. Get started with the system administration guide.

Also in this release

• .msg and .eml files can be imported directly.
• A new split mode looks for document borders inside each file. Until now every file in a transaction was merged into one stream of pages before borders were found. The new mode also covers the case where a document never spans more than one file, but a single file may hold one or several documents. See split modes.
• Data transformation can write to the transaction activity log, which makes debugging a transformation far easier.
• Fields can be filled from transaction metadata, and empty fields can take a default value.
• Cascade classification. A pre-classification rule can narrow the range of classes the model considers, and a new “always” condition covers the case where no real classification is needed because every document is of a known class.