Custom Document Extractor with Document AI Workbench in the Google Cloud Console

This repository demonstrates how to create, train, and deploy a custom document extractor using Google Cloud's Document AI Workbench. The project involves setting up the environment, defining processor fields, labeling documents, training models, and deploying the trained models. Below is a detailed technical overview of each step involved:

Project Setup:

1- Create a Processor:

In the Google Cloud console, navigate to the Document AI section and create a new processor for custom document extraction. Name the processor (e.g., my-custom-document-extractor), select the nearest region, and opt for Google-managed storage and encryption.

2- Processor Configuration:

i. Define Processor Fields:

Specify the fields to be extracted by the processor. Create fields such as control_number, employees_social_security_number, employer_identification_number, employers_name_address_and_zip_code, federal_income_tax_withheld, social_security_tax_withheld, social_security_wages, and wages_tips_other_compensation.

ii. Field Attributes:

Define attributes like Data Type (e.g., Number, Money, Address) and Occurrence (e.g., Optional multiple, Required multiple).

3- Document Labeling:

i. Upload Sample Document:

Import sample documents from Cloud Storage to the labeling console.

ii. Annotation:

Use tools like bounding box and select text to annotate documents. Confirm suggested labels by the foundation model, and manually label fields that are not identified automatically.

4- Model Training:

i. Build Processor Version:

Create a processor version using the pretrained foundation model for initial extraction. Name the version (e.g., w2-foundation-model).

ii. Auto-Labeling with Generative AI:

Import additional documents, split data for training and testing, and use the foundation model to auto-label documents. Verify and manually correct auto-labeled documents.

5- Training Custom Model-Based Processor:

i. Data Requirements:

Ensure at least 10 training and 10 test instances for each field.

ii. Training Setup:

Name the processor version (e.g., w2-custom-model), and select the Model based training method. Start training, which might take several hours.

6- Deployment:

i. Deploy Processor Version:

Deploy the trained processor version. Set it as the Default version or use the version ID for document processing.

7- Evaluation and Testing:

i. Evaluate Processor:

Test the processor version using new documents not involved in training. Assess performance using metrics like F1 score, precision, and recall.

8- Using the Processor:

i. Manage Processor Versions:

Handle different processor versions similarly to other processors in Document AI.

ii. API Integration:

Use Document AI API for online or batch processing. Follow code samples for sending processing requests and handling responses.

Key Insights and Summary

1- Automated Document Processing:

The custom document extractor significantly reduces manual data entry and increases accuracy by automating the extraction of structured data from unstructured documents.

2- Flexible Configuration:

Users can define specific fields and attributes tailored to their document types, ensuring high relevance and precision.

3- Efficient Annotation:

Leveraging foundation models and generative AI for auto-labeling streamlines the annotation process, saving time and improving labeling efficiency.

4- Robust Training and Evaluation:

Training custom models with sufficient data and rigorous evaluation ensures reliable performance in diverse document processing scenarios.

5- Scalable Deployment:

The solution supports scalable deployment and management of multiple processor versions, catering to evolving business needs and document types.

6- API Integration:

Seamless integration with the Document AI API allows for flexible and scalable document processing in various applications and workflows.

This technical description and summary provide a comprehensive guide to building a custom document extractor, automating structured data extraction from various document types using Google Cloud's Document AI Workbench.