Benchmarking document information localization with Amazon Nova

This article details a benchmarking study on document information localization using Amazon Nova, a multimodal large language model available through Amazon Bedrock. The research focuses on precisely identifying and extracting specific information from complex documents.

• Evaluated document localization using the FATURA dataset of 10,000 invoices with 50 different templates
• Developed two localization strategies: image dimension and scaled coordinate approaches
• Amazon Nova Pro achieved a mean Average Precision (mAP) of 0.8305
• Demonstrated strong performance across various document layouts, with consistent accuracy above 0.80
• Showed particular strength in locating structured fields like invoice numbers and dates

The study highlights how multimodal AI models can simplify document processing by reducing the need for complex computer vision architectures and extensive training data.