AI Reads Handwritten Indian Docs Flawlessly

Sarvam AI has unveiled Sarvam Vision 2.1, an enhanced AI model designed to process a wide array of documents, including handwritten records, intricate forms, and extensive tables. This innovative tool supports both English and 22 distinct Indian languages, aiming to broaden its applicability in the business sector for document digitization and information extraction. The previous iteration faced challenges with complex documents and scalability concerns. Sarvam AI addressed these issues by training the updated model on a diverse dataset of real and synthetic documents, simultaneously working to reduce operational costs. While specific price reductions were not disclosed, the company is focused on making the technology more accessible. Sarvam AI also released a comprehensive benchmark dataset comprising 6,909 document samples, with a focus on Indian languages. On this test, Sarvam Vision 2.1 achieved an overall accuracy score of 87.3%, surpassing a competitor's score of 84.9%. However, the model's performance varied significantly across different languages, with scores above 90% for Hindi, Kannada, and Telugu, but lower for Santali (53.9%) and Kashmiri (54.8%). The company acknowledges that benchmark results do not always equate to real-world utility.