The field extraction results include a confidence score for each extracted field. This score indicates how certain is about the accuracy of the extracted data. Having a confidence score allows you to create logic to route fields with low-confidence scores to human reviewers before sending data to downstream systems. For example, you can write a script that sends an extracted field to a human reviewer if the confidence score is lower than a set threshold. The higher the confidence score, the more confident is that the prediction is accurate.
The confidence score feature is experimental and still in development, and may not return accurate results.

​
Availability

  • API: Confidence scores are included in extraction results starting July 21, 2025.
  • Library: Use agentic-doc library v0.3.1 or later to get confidence scores in extraction results.
  • Playground: The extraction results in the Playground do not include confidence scores.

​
Confidence Scores in Output

The API returns a confidence property for each extracted field within the data.extraction_metadata object. The sample extraction_metadata output below shows how the confidence property displays after extraction.
Confidence Score in Output
{
   'patient':{
      'name':{
         'value':'John Doe',
         'chunk_references':[
            '0a75d377-a435-49d0-b987-ee77e67e746c'
         ],
         'confidence':0.96
      },
      'address':{
         'value':'123 Main Street',
         'chunk_references':[
            '1a753ba4-e375-4e70-aaf5-db3834a7d6a7'
         ],
         'confidence':0.83
      }
   }
}

​
Example Workflow: Get Confidence Scores During Field Extraction

This example walks you through how to access and identify the confidence property during field extraction with .
  1. Download this PDF and save it to a local directory: Sample Bank Statement.
  2. Copy the Python script below and save it to a local directory:
    Sample Python Script for Field Extraction
        
from __future__ import annotations

from pydantic import BaseModel, Field
from agentic_doc.parse import parse


class SampleExtractionSchema(BaseModel):
    accountHolder: str = Field(
        ...,
        description='The full name of the person who holds the bank account.',
        title='Account Holder Name',
    )
    accountNumber: str = Field(
        ...,
        description='The bank account number associated with the account holder.',
        title='Bank Account Number',
    )

# Parse a file and extract the fields
results = parse("estatement.pdf", extraction_model=SampleExtractionSchema)
fields = results[0].extraction

# Return the value of the extracted fields
print("Extracted Fields:")
print(fields)    

# Return the value of the extracted field metadata
print("\nExtraction Metadata:")
print(results[0].extraction_metadata)
  3. Install the agentic-doc library.
  4. Get your API key and set it. For more information, go to API Key.
  5. Run the Python script.
  6. Check that the output is similar to the following:
    Confidence Score in Output
    Extracted Fields:
accountHolder='SUSAN SAMPLE' accountNumber='02782-5094431'

Extraction Metadata:
accountHolder=MetadataType[str](value='SUSAN SAMPLE', chunk_references=['d04c75fd-5b1e-4aab-a859-b4aff17305cc'], confidence=0.9999998063873693) accountNumber=MetadataType[str](value='02782-5094431', chunk_references=['5c06add1-bb8e-4143-9424-e95b5bbaeee5'], confidence=0.9718049806957251)

​
Null Confidence Scores

Because the confidence score feature is experimental and still in development, there are certain situations where scores are not available. The confidence score value will be null in the following situations:
  • Tables: Data extracted from tables will have a null confidence score.
  • Changes to formatting: Fields with custom formatting applied during extraction will have a null confidence score. For example, reformatting a date from “DD-MM-YYYY” to “MM-DD-YYYY” results in a null score.