Skip to content

Before anyone asks a question

Document intelligence that keeps the original in view.

Marella turns your files into searchable passages, and shows you what was read, where it came from and what failed.

Test difficult documentsSee the processing layers
Example document · not customer data

You should be able to trace an answer back to the page it came from

Review required
Northstar ComponentsSupplier review · 2026

3. Delivery and continuity

The supplier maintains two approved distribution routes. The secondary route is subject to annual continuity review.

RouteReview status
PrimaryApproved
SecondaryReview due

¶ 13.4 Continuity evidence must be renewed before the next supplier assessment.

Supporting schedules and referenced control owners follow on the next page.

One file, five separate jobs

“Uploaded” does not mean readable.

A file can arrive successfully and still fail later.

  1. 01

    Receive

    Store the file with source identity and a visible processing state.

    Test · Check duplicate, update and deletion behaviour.

  2. 02

    Read

    Extract digital text or use an approved OCR path for image-based pages.

    Test · Measure poor scans, handwriting, layout and character errors separately.

  3. 03

    Segment

    Create passages with metadata and source coordinates where the format supports them.

    Test · Inspect reading order, page boundaries, tables and missing sections.

  4. 04

    Structure

    Extract entities, concepts or fields only where the target workflow needs them.

    Test · Score each field or relationship; a filled-in screen is not proof it is right.

  5. 05

    Reconcile

    Compare the source manifest with processed output, failures and review states.

    Test · Assign retry, exception and human-review ownership before production use.

Fit before features

Sometimes the right answer is a parser, form or database query.

Use a fixed rule

The fields sit in the same place every time.

Use a fixed parser, form or structured query if the fields are always in the same place and the answer has to be exact.

  • Known field locations
  • Deterministic validation
  • Exact record lookup
  • No synthesis required
Consider document intelligence

The layout changes and the evidence is spread across pages.

Use document intelligence when the layout varies, the evidence is spread across pages and someone has to review it.

  • Mixed reports and agreements
  • Digital and scanned material
  • Cross-document questions
  • Named review owner

Source manifest

Reconcile what arrived with what became usable.

A full progress bar does not mean every file was read. Record what is missing, what only half-worked and what needs a person.

What the parser can and cannot do
Digital policyPassedPage + passageReady for retrieval test
Scanned appendixOCR reviewPage onlySample characters and numbers
Wide tablePartialCell uncertainReconcile expected fields
Password-protected fileFailedUnavailableResolve or exclude explicitly

What can still go wrong

Good document AI shows you what it could not read.

  • Poor OCR propagates into retrieval and extraction.
  • Page coordinates are not available for every parsed format.
  • Extraction quality varies by document and field.
  • Document AI output remains subject to human review where the workflow requires it.

Document intelligence FAQ

Is document intelligence the same as OCR?

No. OCR converts image text into characters. Document intelligence may add parsing, extraction, retrieval, synthesis and workflow controls, while still depending on OCR for scans.

Which formats are supported?

We make no universal format claim, because parsing behaviour varies with file structure, scan quality and configuration rather than with the file extension. Ask for a format-and-quality matrix covering the release you are being offered, then run it against your own difficult files.

How should failed documents be handled?

Reconcile the source manifest with media-library states and define retry, exception and reviewer ownership before production use.

Bring the scans, the wide tables and the long files with awkward layouts.

Tell us the file types and what you need out of them, and we will agree a secure way to send them.

Scope a document test
Where the files come fromOneDriveAmazon S3Email