The /form-processing command allows you to extract form fields and their values using a machine learning model specifically trained for this purpose. Over the last year, LLM models like Claude and OpenAI have eclipsed custom models like this. But every now and then, a custom trained model like this may have greater accuracy. So if you’re struggling with a LLM to do document processing on scanned or PDF forms, give this a try.

​

Basic Usage

Process a form by either uploading a file or providing a URL:

/form-processing
File: [Upload your form file]

or

/form-processing
URL: [URL to your form file]

​

Input Formats

Supported file types include:

• PDF documents
• Scanned images (PNG, JPEG)
• Digital forms

​

Output Format

The command returns a JSON object containing:

1. Field Extractions:

   • Field names and their extracted values
   • Confidence scores for each extraction
   • Field types (e.g., checkbox status, text fields)

2. Metadata:

   • Processing timestamp
   • File information
   • Model version

Example output:

{
    "json": {
        "PHARMACY": {
            "fieldValue": "filled_checkbox",
            "confidence": 0.9997177124023438
        },
        "Member ID Number": {
            "fieldValue": "871381692",
            "confidence": 0.9334482550621033
        },
        "Date of Birth:": {
            "fieldValue": "20-MARCH-1960",
            "confidence": 0.9115520119667053
        }
    },
    "table": []
}

​

Understanding the Results

​

Field Values

• Text fields: Contains the extracted text
• Checkboxes: Reports as either “filled_checkbox” or “unfilled_checkbox”
• Empty fields: May return ”-” or be omitted

​

Confidence Scores

• Range from 0 to 1 (0% to 100% confidence)
• Higher scores indicate greater confidence in the extraction
• Generally:
  • greater than 0.9: Very high confidence
  • 0.7-0.9: Good confidence
  • less than 0.7: Lower confidence, may need review

The Form Processing command provides a reliable way to extract data from structured forms. While newer LLM models have broader capabilities, this specialized tool can still outperform them for specific form processing tasks.