Transforming Efficiency: How GenAI Revolutionized Document Data Extraction

Customizing GenAI Solutions for GAEA Global

In an age of fast-paced technological advancements, countless companies are intrigued by the potential of GenAI and seek to incorporate it into their operations. Yet, many face uncertainty on where to begin.

This case study illuminates how we supported GAEA Global, a prominent provider of supply chain solutions, in pinpointing the optimal GenAI solution tailored to their specific industry requirements.

What we anticipated along the journey.

Maintenance logs and audit reports are currently paper-based and exist in various formats. These reports are manually filled out and stored for future use. Digitizing this valuable data presents numerous business opportunities, including enhancing operational efficiency. However, due to the diversity of templates used, developing a traditional system is both time-consuming and costly.

How we created a hybrid approach utilizing ML and Generative AI

Our goal was to address this issue by leveraging cutting-edge techniques powered by LLMs / ML models to achieve the following:

• Extract data from various reports and transform it into a format suitable for downstream consumption and processing.

• Identify the structure (fields, tables) of the report and generate a schema that facilitates the creation of digital forms for collecting data electronically.

Numerous ML-powered document data extraction tools are available, with some utilizing LLM-based AI models. We embarked on a thorough exploration of the landscape to identify the most promising options and conducted experiments with the top contenders over a span of two weeks to evaluate their capabilities.

Our findings revealed that while the form parser effectively extracted most of the keys from the documents, the custom extractor, powered by generative AI, excelled at extracting the values when the keys were pre-configured. To optimize our approach, we adopted a hybrid strategy, integrating both processors through their APIs and passing our documents through this combined solution.

To our satisfaction, this hybrid approach outperformed using the processors individually.

Striking the Balance Between Innovation and Accuracy in Generative AI

As the landscape of Generative AI rapidly evolves and our understanding of LLMs continues to expand exponentially, it's important to recognize that these findings may have a short shelf life, reflecting the accelerated pace of innovation in the AI-driven world.

While zero-shot training methods (such as the Form Parser and ChatGPT) offer expedited deployment, they provide limited opportunity for fine-tuning results to enhance accuracy. This approach may not be suitable for scenarios where data accuracy is paramount within critical business workflows.

In contrast, dedicating time to customize or fine-tune data extraction requirements (utilizing the Custom Extractor or Hybrid Approach) yields higher and more predictable accuracy levels. Despite the initial effort required to define the desired keys for extraction across a given set of documents, this approach offers greater control and reliability in data extraction.