DOCUMENT READING
Reading what was dropped in.
The content of a PDF or an image becomes text your agents can read, within a deliberately narrow perimeter.
What gets read, and in which formats
Reading covers PDFs and images: PNG, JPEG, WebP. Office files are not part of it. A spreadsheet or a word processor file dropped on a record is stored, previewed and downloaded normally, but its content is not extracted through this path.
What comes out is text, page by page, put back in the order of the document. It is not a filing operation: reading produces text, not filled-in fields.
An extension point, not a black box
The engine defines a reading contract and plugs the service of your choice into it. One implementation ships with the platform; an installation that prefers its own service swaps it out without touching anything else.
The direct consequence: as long as no service is configured for your workspace, reading does not run at all. It is an installation setting, not a checkbox in a settings screen, and it is the first thing to have checked when an agent says it could not read a document.
Processing is asynchronous, and it can fail
Uploading a file queues a job rather than making the person who uploaded it wait. The job is retried on a transient failure, then the file stays marked as unread. Each file therefore carries its own state: queued, running, read, failed.
That state is not surfaced by any screen today. This is a gap, not a subtlety: in practice it is the agent that tells you it could not read a document. An agent can also trigger the reading of a document that never went through it.
What reading does not do
It does not fill in a record's fields. There is no structured extraction going after an IBAN or a due date to write it on your behalf: that feature existed and was removed.
The extracted text does not feed your workspace search either. It is readable by an agent, on a document it has located first. Searching for a phrase inside the content of all your files from the search bar is not something the platform can do today.
How it works
A client sends a scan of their amendment. You want an agent to be able to refer to it while reviewing the file.
The file is dropped on the record, like any other piece.
If a reading service is configured for the workspace, a background job is queued.
The content is extracted page by page, as text.
The text is stored on the file, along with the state of the reading.
An agent opening the document receives that text, and knows if reading failed.
On a transient failure the job runs again; otherwise the file stays marked unread.
The other pieces of this chapter
Agentic drive
Every record carries its files, and your agents read them.
Bulk import
Exports, databases and archives become typed records.
Inbox
Your exchanges surface on the records they concern.
Meetings
Recorded, transcribed, read back from the record.
Calendar
Your events next to your records, read-only.