Loading

Document and Image Intelligence

Document and image intelligence is the platform reading content it cannot otherwise interpret: the words inside uploaded documents, the text in a scanned page, and, through the AI gateway, a model reading a stored document or a picture. This page sets out what exists today, what each part does, and what is not available, so you can plan around the real capabilities.

Where to find it

Architect Panel → Automation:

  • Tasks — Document Text Extraction, which reads documents in the background every 15 minutes

Architect Panel → Configuration:

  • AI Settings — the ocr role, when scans are read by an AI model

Architect Panel → Activity:

  • AI Usage — the cost of each scan read by AI, under the feature ocr

Architect Panel → Security:

  • Classification Levels — Allow Export, which decides whether a document may be read off-site

What exists

  • Text extraction for search. PDFs, Word and OpenDocument files, spreadsheets, presentations, and plain text, HTML and e-mail are read into the document index so their contents can be searched.
  • Reading scans (OCR). A scanned image, or a PDF that is only a picture of text, needs optical character recognition. Your hosting administrator chooses the reader in the server's configuration: Tesseract on your own server, AWS Textract, or an AI model through the gateway. It is off until chosen.
  • AI reading of scans. With the AI reader, each scan is sent to the model on the ocr role and asked for a transcription and nothing else. See the next article.
  • Stored documents as AI input. A custom feature or an MCP client can give a model a stored document, under the export rules. See AI Gateway.

What is not available

  • Language detection and image labelling have no screen, setting or automatic use on the platform. They exist only as building blocks a developer can use in custom code.
  • Pulling fields out of an arbitrary document, such as the total on any invoice, is not a configurable feature. One installation-specific extractor reads supplier bills arriving from Xero, and it is not configured from any panel screen.

Not the same as Data Extraction

Data Extraction ingests structured files such as CSV, JSON and XML, where a value sits at a known place. Everything on this page reads unstructured content, where nothing is at a known position and the result is a best reading. If your source is a spreadsheet export, use Data Extraction.

Extraction runs in the background

Reading a document can take seconds to tens of seconds, so uploads are queued and the Document Text Extraction task works through the queue every 15 minutes. It ships disabled: until it is switched on, nothing queued is ever read, nothing new is searchable, and only Claude can be sent a stored PDF.

  1. In Tasks, switch on Document Text Extraction.
  2. Ask your hosting administrator which OCR reader, if any, is configured, and whether it reads files on your server or elsewhere.
  3. Upload one document of each type you expect, including a scan, and search for a phrase from each after the next run.

Results are readings, not facts

Text from a clean, machine-made PDF is exact. Text read from a scan or photograph is approximately right: good enough to find a document, not good enough to rely on for a figure or a reference number without someone checking it against the original. Quality in is quality out: a straight, well-lit scan reads far better than a photograph taken at an angle.

What goes wrong, and how to tell

  • Nothing new is searchable: the extraction task is off.
  • Scans never become searchable: no OCR reader is configured, or the reader is off-site and the document's classification forbids export. The latter is recorded as a failure naming the reason.
  • A scan is "Not read yet" with a sentence about AI credit: the AI reader was refused because the tenant's AI credit is used up. It is not re-read automatically after a top-up.

Worked example

An organisation with a library of scanned correspondence switches on Document Text Extraction and asks its hosting administrator to configure Tesseract, because some of the archive is classified and must not leave. Typed PDFs become searchable within the hour; scans follow as the queue is worked through. Nobody expects the platform to detect the language of letters, because the team checked this page first.

Recommendations

  • Switch on the extraction task, or nothing is ever read.
  • Choose the OCR reader with your classification in mind: only a local reader can read everything.
  • Check a sample of each document type after the first run.
  • Treat text from scans as a strong suggestion, checked against the original where it matters.