Skip to main content
Roberty Studio
IDPBuilder

Extraction

Choose how the content comes out of the document in an IDP template: Roberty Cloud OCR, with text, tables, forms and block positions, or LLM OCR, guided by an instruction.

Extraction

Extraction gets the content out of the document after pre-processing. Each template uses exactly one extraction method, picked from the stage's cards. A new template starts with Roberty Cloud OCR and the Extract tables option on.

MethodWhat it does
Roberty Cloud OCRManaged OCR with support for tables, forms and text block positions.
LLM OCRExtraction guided by a natural-language instruction, carried out by the template's model.
info

Each method depends on a feature enabled on the workspace: Roberty Cloud OCR needs the feature of the same name, and LLM OCR needs the generative AI feature. Without the feature, the method's card shows as locked. Contact your account manager to have the feature turned on.

Roberty Cloud OCR​

Roberty Cloud OCR always returns the document's text, line by line. The stage's options add structured data to the result:

OptionDefaultWhat it adds
Extract tablesOntables: every table found, with its rows and columns as lists of cells.
Extract formsOffforms: key and value pairs, such as "Tax ID: …". Checkboxes show up as [X] (checked) or [ ] (unchecked).
Extract text block positionsOffblocks: each line of text with its confidence, from 0 to 1, and the coordinates of the box around it on the page, also from 0 to 1.

With all three options off, the result is just the text. Each option you turn on increases the workspace's OCR usage: every processed page counts as one request, plus one more per option turned on.

Pages larger than 5 MB are recompressed automatically before they are read.

In documents with several pages, Roberty Cloud OCR reads page by page and merges the results: each page's text is marked with === Página N de M ===, and tables, forms and blocks carry the number of the page they came from. See Document and pages.

LLM OCR​

LLM OCR sends the page images to the model set in Model, along with your Instruction: say what to extract and in which format, for example "Extract the invoice number, issue date and total amount.". All selected pages go to the model in a single call, so it reads the whole document at once.

With the instruction empty, the model transcribes all readable text in the document, keeping the reading order.

The result of LLM OCR is always text: the model's answer, with no separate tables, forms or blocks. If you ask for a JSON in the instruction, it reaches post-processing as text.

What goes on to post-processing​

The extraction result is the starting point of post-processing:

  • when there is only text — LLM OCR, or Roberty Cloud OCR with all three options off — the first step receives the text as a string;
  • when there is structured data, the first step receives an object { text, tables, forms, blocks }, with only the parts that were extracted.

Without post-processing, this result is returned as is.

Which method to choose​

SituationMethod
Documents with tables, form fields or checkboxes, or when you need the position of the text on the page.Roberty Cloud OCR
When you want to describe in natural language what to extract and in which format, and the model should read all pages together.LLM OCR

A common setup pairs Roberty Cloud OCR with an LLM or code step in post-processing that turns the text and tables into the final JSON.