Quality control the data extractions in a document type by writing validations using JsonLogic. Test extracted fields using Boolean, logic, numeric, array, string, and other operations. Then write your own logic based on the validations, for example:

• pass a document extraction automatically through your pipeline if there are no errors and 10% of warning validations fail
• flag a document extraction for human review if 5% of error validations fail

Create validations

Sensible app

To create validations in the Sensible app:

1. Click the document type.
2. Click Create validation.
3. Enter the parameters for the validation.
4. Click Create.

Parameters

A validation has the following parameters:

Examples

Imagine you have a document type "sales_quotes" with configs for

• company_A
• company_B
• company_C

You test sales quote extractions from all the companies with the following validations:

Validation 1

• Description: The quoted rate value isn't null
• Severity: error
• Condition:{"exists":[{"var":"quote_rate.value"}]}

Notes: Tests that a field (quote_rate) isn't null using the Sensible exists operation.

Validation 2

• Description: The quote duration is a round number
• Severity: warning
• Condition:{"==":[{"%":[{"var":"quote_duration.value"},2]},0]}

Notes: Retrieves the value of an extracted quote_duration field using the JsonLogic var operation, then uses the JsonLogic modulo operation (%) to divide the rate by 2 and passes the test if the remainder equals ("==") 0.

Validation 3

• Description: Broker's email is in [email protected] format
• Severity: warning
• Prerequisite fields: ["broker\\.email"]
• Condition: {"match":[{"var":"broker\\.email.value"},"^\\S+\\@\\S+$"]}

Notes: If broker.email isn't null, then uses a Sensible operation (match) to test that the email matches a regular expression. If broker.email is null, skips this condition.

Validation 4

• Description: The zip code is valid for USA or CA

• Severity: warning

• Condition:

{"or":[
  {"and":[
    {"==":[{"var":"country.value"},"US"]},
    {"match":[{ "var":"zip_code.value" },"^[0-9]{5}$"]}]},
  {"and":[
    {"==":[{"var":"country.value"},"CA"]},
    {"match":[{"var":"zip_code.value"},"^[A-Z][0-9][A-Z] [0-9][A-Z][0-9]$"]}]}
]} 

Notes: Tests that the zip_code is a 5-digit number if the country field equals USA, or 6 alphanumeric characters if the country field equals Canada. Uses a Sensible operation (match) to test regular expressions.

Validations output

The following extraction excerpt shows validating test data with the preceding conditions:

• Validation 3 Sensible skips the broker email because the prerequisite field broker.email is null
• Validation 4 (zip code) fails because zip_code is 17 digits

{
    "parsed_document": {
        "quote_rate": {
            "source": "$800",
            "value": 800,
            "unit": "$",
            "type": "currency"
        },
        "quote_duration": {
            "type": "number",
            "value": "6"
        },
        "broker.email": null,
        "country": {
            "type": "string",
            "value": "USA"
        },
        "zip_code": {
            "type": "number",
            "value": "12345678901234456"
        },
        "validations": [{
            "description": "Zip code must be valid",
            "severity": "warning"
        }, {
            "description": "Broker's email is in [email protected] format",
            "severity": "skipped",
            "message": "Missing prerequisites: broker.email"
        }],
        "validation_summary": {
            "fields": 5,
            "fields_present": 4,
            "errors": 0,
            "warnings": 1,
            "skipped": 1
        }
    }
}

Notes

Why does Sensible use validation tests rather than confidence intervals? Sensible's extractions are largely deterministic. With the exception of OCR-dependent output, a Sensible config always returns the same output for a given PDF input. Given this, deterministic validation tests are more useful than confidence intervals as measures of extraction quality.