You can extract specified fields from a document when parsing it with the agentic-doc library.

There are two methods for running extraction with the library:

  • Define the fields that you want to extract using Pydantic models directly in your code. Then pass that definition to the extraction_model parameter when running the parse function.
  • Input the schema in the extraction_schema parameter as a dict and receive the answer as dict (JSON).

​
Sample Script: Extract Fields (extraction_model)

Run this script to extract fields from a local file and return the value of one of the extracted fields. In this scenario, the schema is passed in the extraction_model parameter.

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

# Define the fields you want to extract
class ExtractedFields(BaseModel):
    employee_name: str = Field(description="the full name of the employee")
    employee_ssn: str = Field(description="the social security number of the employee")
    gross_pay: float = Field(description="the gross pay of the employee")
    employee_address: str = Field(description="the address of the employee")

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

# Return the value of one of the extracted fields
print(fields.employee_name)

​
Sample Script: Extract Nested Subfields

You can extract nested subfields from documents. For example, let’s say you have a medical form that has two “name” fields: one for the Patient and one for their Emergency Contact. You can extract both fields by running the script below.

The script extracts the nested subfields from a local file and returns their values.

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

# Define the main extraction model with nested subfields
class MedicalField(BaseModel):
    patient: Patient = Field(description="All information pertaining directly to the patient")
    emergency_content: EmergencyContact = Field(description="All information pertaining directly to the patient's emergency contact")

# Define the nested field classes
class Patient(BaseModel):
    name: str = Field(description="Patient name")

class EmergencyContact(BaseModel):
    name: str = Field(description="Emergency contact name")

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

# Return the values of the extracted nested fields
print(fields.patient.name)
print(fields.emergency_contact.name)

​
Extraction Output

When you extract fields from a document using the agentic-doc library, these fields are returned:

​
extracted_schema

This is the same type as the BaseModel that is passed in as the argument to extraction_model.

​
extraction_metadata

This is the same type as the BaseModel that is passed in as the argument to extraction_model.

It has the same nesting structure as the BaseModel except that each field is replaced with a dictionary. This dictionary uses this form:

{"chunk_references": ["chunk-id-1", chunk-id-2", ...]}

Here is an example of a returned extraction_metadata field:

extraction_metadata = {
    "employee_name": {
        "chunk_references": [
            "72ba3cca-01e5-407b-9fc4-81f54f9f0c51"
        ]
    },
    "gross_pay": {
        "chunk_references": [
            "5b8865b9-1a81-46df-bcf7-0bdbed9130dc"
        ]
    }
}

If you have a more deeply-nested extraction_model (like in this medical form example), then you may see something like this:

extraction_metadata = {
    "patient": {
        "name": {
            "chunk_references": [
                "72ba3cca-01e5-407b-9fc4-81f54f9f0c51"
            ]
        }
    },
    "emergency_contact": {
        "name": {
            "chunk_references": [
                "5b8865b9-1a81-46df-bcf7-0bdbed9130dc"
            ]
        }
    }
}

​
extraction_error

If the extraction function encountered an error, the extraction_error field will describe the issue. If there are no errors, the extraction_error field returns None.