Beyond OCR: Achieving 98% billing accuracy with GroundX and Red Hat OpenShift AI

In the enterprise world, document extraction has been stuck in a 20-year rut. For decades, companies have relied on a fragile assembly line: OCR to convert pixels to text, templates to find fields, and a team of human reviewers to fix the inevitable 30% error rate.

When a utility provider changes a font or a customer submits a photo of a crumpled bill, these legacy systems break. To solve this, we need to move away from rigid string searches and toward semantic understanding.

The architecture

Figure 1 illustrates the components of an AI-based solution, using components from Red Hat and Valantor, that can markedly improve the quality of the document extraction process. 

• GroundX provides the "brain," which achieves up to 98% accuracy when processing documents
• Red Hat OpenShift AI provides the "nerves," serving the models and managing the agent loops.
• Red Hat OpenShift provides the "body," so the entire system is security-forward, compliant, and ready for production.

We’ll discuss each of these components in turn.

Figure 1: The architecture of our document extraction solution.

The foundation: Red Hat OpenShift

Before you can deploy sophisticated AI, you need a resilient infrastructure. Red Hat OpenShift is the enterprise-grade container platform that serves as the bedrock of this solution, with Kubernetes as one of its key underlying technologies.

In a billing extraction workflow, Red Hat OpenShift provides the hybrid cloud substrate. Whether your documents live in an on-premise vault or a public cloud bucket, Red Hat OpenShift keeps your processing pipeline portable and auto-scales to meet peak demand—during end-of-month billing cycles, for instance.

The engine: Red Hat OpenShift AI

If Red Hat OpenShift is the bedrock, Red Hat OpenShift AI is the specialized engine. This platform provides the tools to manage the model serving and monitoring.

In this architecture, OpenShift AI acts as the MLOps orchestrator. It provides the environment where specialized LLMs and vision models are served, allowing developers to manage the agentic workflows required for complex document interpretation without worrying about the underlying hardware complexities.

The intelligence: GroundX from Valantor

At the top of the stack is GroundX, an AI-native platform that collapses data ingestion, storage, search, and evaluation into a single, unified pipeline. GroundX merges a vision model trained on nearly 1 million page extractions of enterprise documents with a visual language model and an agent pipeline. The result is an engine that interprets files as a human would—by looking at layout, tables, graphics and text simultaneously.

GroundX goes beyond basic text extraction by recognizing structural elements like tables, charts, and paragraphs. It then uses this understanding to tailor its output, generating data in the specific format best suited for your downstream task, whether that is search retrieval, summarization, or structured JSON extraction.

Accelerated data extraction for AI agents

While manual OCR apps generate flat, unstructured text that enterprise tools struggle to reliably interpret, GroundX translates complex documents into structured, context-aware data tailored specifically for machine consumption. In an agentic AI environment, autonomous agents can easily consume this service to understand spatial layouts, tables, and concepts without needing rigid templates. This deep semantic understanding eliminates manual bottlenecks, empowering your AI workflows to extract and act upon information with high accuracy at mass scale. The shift from OCR to GroundX’s agentic architecture offers 3 big advantages:

• 98% accuracy with no templates: Unlike legacy systems that require a new template for every bill format, GroundX is able to understand and process new, previously unseen document formats without needing to be explicitly trained on them or given a specific template beforehand. For instance, it understands what "Total Balance" means on a bill, regardless of where it appears on the page.
• Contextual awareness: GroundX performs a context pass that compares extracted data against surrounding content and document-level summaries. This means that a table at the bottom of a page will be interpreted correctly based on the text at the top, for example.
• Handling messy data at scale: The vision-language stack can process mobile phone photos, crumpled paper, and challenging lighting, conditions that can typically break standard OCR. Removing manual steps and collapsing multiple steps into a single unified process to be consumed by AI agents enables data extraction at scale.

Enterprise AI is only as good as its deployment. In this integration, Red Hat OpenShift and OpenShift AI provide the infrastructure that supports GroundX’s AI capabilities. Together, they form a solution that helps eliminate the glue code typically required to stitch together OCR strings and regex matches, returning clean, typed JSON that is ready for your database.

Ready to automate?

The days of manual review and brittle templates are over. You can now transition from 70% accuracy to 98% by modernizing your document stack.

Try the AI Quickstart: Get up and running with Red Hat OpenShift, OpenShift AI, and GroundX in a pre-configured environment. Test your own messy documents and see the difference that semantic understanding makes.

Explore the GroundX AI quickstart and the GroundX documentation.