What do I need to run form extraction?

APEX Document Management (Oracle Database 19c or later, Oracle APEX 24.2 or later) with the AI Pack and its install prerequisites, and access to a language model. The AI Pack calls models through UC AI, an open-source PL/SQL library. OpenAI, Anthropic, Google, OCI Generative AI, Mistral and Ollama work. The API key is stored in an APEX web credential. For images you need a vision model, and for scanned PDFs an OCR service.

Do I have to write PL/SQL?

Not for the rules. The configuration is JSON, and the checks are SQL conditions. You can edit all of it in the form extraction manager, an APEX region plug-in that you add to pages of your application. Without the manager, you save the same JSON with calls of the PL/SQL API. PL/SQL is also where you connect your own systems: the AFTER_FORM_EXTRACTION hook, and your own review pages for users who are not administrators.

What stops a wrong value from reaching my ERP system?

Checks. The JSON schema makes fields required and sets their types. A check is a SQL condition on the extracted values, for example that the net amount plus VAT equals the total, or that a quantity is positive. Your hook can also refuse a result, for example when a purchase order does not exist in your database. A result that fails is not saved on the document. It goes to review and waits for a person. Checks catch values that do not fit together. They cannot prove that every value is right: a name or a single check box has nothing to compare with. Measure such fields and decide whether a review is the safer choice.

Does the model decide which form a document is?

Not for documents with text. Each form has markers, regular expressions that must all match the text of the document. Exactly one form must match, otherwise the document goes to review. Only an image without text is classified by the vision model, because there is no text to match.

Can it read scans and images?

Yes. An image goes to the vision model of the profile, which picks the form and reads the values. A scanned PDF without a text layer needs OCR: an OCR service (Mistral, OCI or Ollama) turns it into text first. Without OCR, a scanned PDF goes to review. In our test, four simulated scans, supplied as PNG images, went to the vision model. With the final configuration, every extraction of them was correct. In an earlier run, one timesheet scan was read incorrectly. Scanned PDFs were not part of the test.

Where are my documents sent?

Documents and results stay in your Oracle database. For an extraction, ADM sends the text of the document, or the file itself for forms that read the file, to the model provider that you configure. With an Ollama server in your own network, the document stays there.

What does an extraction cost?

The price of the model tokens at your provider. In our test, one run of all 31 documents with gpt-5.6-luna used about 45,000 prompt tokens and 16,000 completion tokens. On eight PDFs, reading the file instead of the text used about three times the prompt tokens. The AI Pack is included with both APEX Document Management licenses and is optional to install. Contact us for licensing details.

AI Pack for APEX Document Management

Turn invoices and forms into checked data.
Run from your Oracle database.

Q: How do I know it is good enough for my documents?

Measure it. By default, approved results go into a test set with their right values. measure extracts every document of the test set again and reports, for each document, whether it was right, caught by a check, a false alarm, or wrong and saved. It also records the tokens and the seconds, and it can run the same test set with another model, so you can compare quality and cost.

A language model reads the fields of your PDFs and images. Checks that you write in SQL decide which values are saved as metadata on the document, ready for your reports or your ERP. Everything else waits in a review queue for a person.

Request a test with your documents Start the tutorial

Or read the form extraction guide.

The upper half of an invoice from Harbour Office Supplies GmbH to Brightline Consulting AG, number INV-2026-00412, with three lines and a total of EUR 557.63 — A sample invoice from the tutorial, and what a small, low-cost model extracted from it.

What it reads

PDFs, images and scans, and what each one needs

PDFs with text

ADM reads the text with Oracle Text, the same text that its search uses. Other files that Oracle Text reads, such as Word documents, work the same way.

Needs: A language model.

Images

PNG, JPEG and other images go to a vision model, which recognizes the form and reads the values.

Needs: A vision model.

Scanned PDFs

A PDF without a text layer has nothing to read. An OCR service turns it into text first. Without OCR, the document goes to review.

Needs: An OCR service.

Grids and check boxes

Their layout carries the meaning, and the text loses it. Set such a form to send the PDF itself to a vision model.

Needs: A vision model that accepts PDF files.

How it works

The model reads. Your rules decide.

Every document goes through the same five steps. Which form a document is, and whether its values are good enough, is decided by configuration that you control.

Read

ADM gets the text of the document, or sends an image to a vision model.
Match

Markers, patterns that you set for each layout, decide which form a document with text is, not the model. For an image without text, the vision model picks the form.
Extract

The model returns exactly the fields of your schema.
Check

Your checks test the values: required, of the right type, and consistent with each other.
Save or review

All checks pass: the values are saved as metadata on the document. Otherwise a person reviews it.

A saved value is an annotation on the document, such as invoice.total_amount = 557.63. Users search and filter on it, your reports read it from a view, and a hook can copy it into your ERP in the same transaction.

Checks

What checks catch, and what they cannot

Required fields and types

The JSON schema makes a field required, a number, or one of a list of values.

Arithmetic

Lines add up to the net amount, net plus VAT is the total, opening balance plus transactions is the closing balance.

Ranges and rules

A quantity is positive, a VAT rate is 7 or 19, a credit note has a negative total.

Lookups in your database, in your hook

Your PL/SQL hook can check values against your tables, for example that the purchase order exists, and refuse the result. The document then goes to review.

What checks cannot do

Checks catch values that do not fit together. They cannot prove that every value is right. A single check box or a free-text remark has no second value to compare with, and nothing to look up. For such fields, measure how often the model is right, and decide whether a review is cheaper than a mistake.

One document type, as JSON

{
  "json_schema": {
    "type": "object",
    "properties": {
      "invoice_number": {"type": "string"},
      "net":   {"type": "number"},
      "vat":   {"type": "number"},
      "total": {"type": "number"}
    },
    "required": ["invoice_number", "total"]
  },
  "annotation_fields": ["invoice_number", "total"],
  "checks": [{
    "name": "net + VAT = total",
    "expr": "abs(net + vat - total) <= 0.01"
  }]
}

The document type says what to read and how to check it. Each supplier layout is a form of that type: a few markers that recognize it, and hints for that layout. A new supplier needs only a new form. Edit all of it in the form extraction manager, an APEX region that shows the text of a document and runs a test extraction before you save, or with the PL/SQL API. The guide covers every setting and the known limits.

Measurement

Measure it. Do not guess.

A model that is right on the three documents you tried can be wrong on the fourth. By default, approved results go into a test set with their right values. One call extracts the whole test set again and tells you, for each document, whether it was right, caught by a check, a false alarm, or wrong and saved. That last number is the one to bring to zero.

It also records the tokens and the seconds, and it can run the same test set with another model. So you compare quality and cost before you switch.

A measurement run in the form extraction manager, with the outcome, the form and the method of each document, and the tokens and seconds of the run — One measurement run in the manager: outcome, form and method for each document, with tokens and seconds. Select the image to enlarge it.

Case study

Not only invoices: a test with five document types

In our sample, checks sent every sign error in a bank statement to review instead of saving it. For timesheets and leave requests, sending the PDF to a vision model kept the layout, and every extraction was correct.

The first page of six sample documents: an English and a German bank statement, a timesheet, a utility bill, a certificate of analysis and a leave request — One sample of each issuer: the bank statement is one document type with an English and a German form. All issuers and values are invented. Select the image to enlarge it.

1. What we tested

31 invented documents from six issuers: bank statements in English and German, timesheets, German utility bills, certificates of analysis and leave requests, plus two letters that are no form. Four are simulated scans, supplied as images.
One model: gpt-5.6-luna, a small, low-cost model. Every extraction was compared with the known correct values.
Tuned on the same documents. The rules, the sign rules for the bank statements, the setting for statements without transactions and the switch to the vision model came from failures in these runs. So the results below show that the final configuration works for these documents. They are not an error rate for new documents.

Not tested: your documents, scanned PDFs (the scans were images), and other model families. Measure on your own forms.

2. Results with the final configuration

Document type	Processing method	Observed result
Certificates of analysis	Extracted PDF text (the scan: vision model)	All 15 extraction attempts correct (5 documents, 3 runs)
Utility bills (German)	Extracted PDF text	All 15 extraction attempts correct (5 documents, 3 runs)
Bank statements (English and German)	Extracted PDF text (the scan: vision model)	All 18 extraction attempts correct (9 documents, 2 runs), after adding sign rules and allowing statements without transactions
Timesheets (grid)	PDF sent to a vision model (the scan: image)	All 10 extraction attempts correct (5 documents, 2 runs)
Leave requests (check boxes)	PDF sent to a vision model (the scan: image)	All 10 extraction attempts correct (5 documents, 2 runs)
Letters (no form)	Form matching on the text	All 6 attempts correctly sent to review (2 letters, 3 runs)

With the final configuration, no attempt saved an incorrect value. Before it, the bank statements already saved none: the balance check sent 2 sign errors to review, and 3 correct statements without transactions went to review because their table was empty. Timesheets and leave requests did save incorrect values when the model read the text: 13 of 30 attempts. That is part 3.

Markers never picked the wrong form. Letters on the same letterhead as a form went to review.
German documents needed one rule for decimal commas, debit and credit signs, and the m³ to kWh conversion.
A provider error saved nothing. When the account ran out of credits, every document became FAILED with a sentence for the reviewer.

3. Why timesheets and leave requests need a vision model

When the model read the extracted text, these two layouts saved incorrect values in 13 of 30 attempts. The values were in the text, but their position was not, and no check can recover information that the text does not contain.

Grids with empty cells

On paper, the 4.0 in the first row is Saturday. In the text, the empty cells are gone: the values follow one another, and nothing says which day the 4.0 belongs to.

A timesheet grid next to the text that Oracle Text returns for it: the values follow each other without the empty cells

Check boxes in two columns

A ticked box is a lone X in the text, between two labels. A single choice has no second value to compare with, so a check cannot tell a wrong one from a right one.

The leave type check boxes of a sick-leave request next to their text, where the X stands on its own line between two labels

One form setting: a vision model reads the PDF

Send the PDF to the vision model

With "model_input": "file", ADM still finds the form with the markers on the text. Then it sends the PDF itself to the vision model of the profile instead of the text. The vision model receives the page layout, including columns and check boxes.

{
  "markers": [
    "Brightline Staffing Ltd",
    "^WEEKLY TIMESHEET$"
  ],
  "model_input": "file"
}

A smaller diagnostic experiment

The same 8 PDFs, text against vision

Input method	Run 1 correct	Run 2 correct
Extracted PDF text	6 of 8	4 of 8
PDF sent to vision model	8 of 8	8 of 8

Sending the PDF increased the median processing time from 2.7 to 4.2 seconds and used about three times as many input tokens. Set it on the forms that need it, not on the whole profile.

The four timesheet and four leave request PDFs, two runs per method, with the same schema, checks and model. These are the documents that showed the problem, so this small sample is no error rate.

Will it work on your documents?

Request a test with your documents

Or build it yourself: start the step-by-step tutorial.

Built into ADM

From the inbox folder to your ERP, in SQL and APEX

Form extraction runs on the documents, folders and permissions of APEX Document Management. The results are rows in your database.

Document metadata

Accepted values are saved as annotations on the document, such as invoice.total_amount = 557.63. Search and filter on them.

A view for reports

adm_my_form_extractions_v returns status, values and issues, only for the documents that the current user may see.

A hook for your ERP

AFTER_FORM_EXTRACTION runs your PL/SQL in the same transaction when a result is accepted.

A background job

Extracts every new document and every new version of an inbox folder. Each one costs at least one model call.

A review queue

Documents that need a person get a review tag. Approve them in the manager, or call approve from your own page.

A manager for administrators

An APEX region: edit the configuration, see the text and the matching markers, test an extraction, run measurements.

Read the results in your own APEX page or report

select document_name, status, form_code, issues
     , json_value(fields, '$.total_amount')
         as total
  from adm_my_form_extractions_v
 where profile_name = 'supplier_invoices';

Model providers

OpenAI, Anthropic, Google, OCI Generative AI, Mistral and Ollama, through the open-source PL/SQL library UC AI. API keys stay in APEX web credentials.

Requirements

Oracle Database 19c or later, Oracle APEX 24.2 or later, APEX Document Management with the AI Pack (included with the ADM license and optional to install; see its install prerequisites), and access to a model provider.

Frequently Asked Questions

Common questions about form extraction in APEX Document Management.

See it on your own documents

Contact us to arrange a test with your documents. We show you the extraction, the checks and a measurement on them. Or build it yourself with the step-by-step tutorial.