Overcoming the Real-World Pitfalls of Google Document AI

This paper discusses the practical and feature gaps that were encountered with Google Document AI in building the AI product at TrialBase platform (ai.trialbase.com), which automates legal document analysis. Results matter because there is an explosion of electronic legal documents that require fast and reliable parsing, which is essential for systems based on LLMs and retrieval-augmented generation. Standard Document AIs seldom work well in practice, even if there are no damaged PDFs, and if a large dataset is being used, wherein the API quota is not hit, and processing costs do not matter. The architecture proposed in this paper is robust, efficient at transforming various documents into structured data. Event-driven microservice architecture with message queues and a PDF sanitization pipeline solves real-world problems by enabling ProcessorPool (multiple processors using synchronous Document AI API to go beyond quota limitation concurrently drastically reducing processing times). Pre-sanitization, coupled with asynchronous batch processing and a custom load balancer, got a tenfold speed increase with enhanced reliability over real-world legal documents. The article is meant to help LegalTech researchers and practitioners, workflow developers, and engineers working on high-performance, reliable Google Cloud-based projects.