The new model focuses on converting unstructured and often difficult-to-read document scans into structured data that can be used directly in business workflows. Sarvam AI said Vision 2.1 brings improvements in complex table processing, form extraction and handwritten text recognition, particularly for Indian languages.
The update builds on the earlier version of Sarvam Vision and addresses challenges commonly faced when AI systems process dense, irregular or handwritten documents.
Sarvam Vision 2.1 focuses on complex documents
Sarvam Vision 2.1 has been developed to understand a range of document formats, including multi-page files, structured forms and complex tables.
One of the key upgrades is its ability to process nested and multi-page tables. Such documents can contain information spread across multiple rows, columns and pages, making them difficult for conventional document-processing systems to interpret accurately.
The model can also extract key-value information from forms. This includes details such as names, dates and financial figures, allowing businesses to convert information from scanned documents into structured data.
Another major capability is handwriting recognition. Vision 2.1 is designed to recognise handwritten notes and annotations in Indic languages, addressing a challenge that becomes more complex when documents contain different regional scripts and handwriting styles.
Improved Indic handwriting recognition
Indian documents frequently combine printed text with handwritten additions, annotations and signatures. Recognising such information accurately is particularly challenging when the handwriting varies between regions and languages.
Sarvam AI said its new model has been trained with a focus on regional handwriting variations and form-extraction use cases.
The company has also introduced training interventions intended to reduce hallucinations and spatial inconsistencies.
Spatial understanding is particularly important for document AI because information is often connected to its location on a page. For example, a number may belong to a particular field or row in a table, while a handwritten note may refer to a specific section of a form.
By improving this understanding, the company aims to make the extracted information more reliable for downstream business applications.
New benchmark covers all 22 official Indian languages
Alongside the launch, Sarvam AI introduced a new Sarvam Indic OCR benchmark to evaluate document-reading performance across India’s 22 official languages.
The evaluation dataset contains 6,909 samples. Of these, 6,609 samples cover the 22 regional languages, while 300 samples serve as English-language baselines.
The dataset includes different types of material, such as newspapers, brochures, textbooks and historical archives. The historical material spans documents dating from 1800 to the present day.
This broad dataset is intended to test OCR performance across different document formats, periods, scripts and language conditions.
Vision 2.1 reports 87.39% word accuracy
According to Sarvam AI, Vision 2.1 achieved an overall word accuracy score of 87.39% on its Indic OCR benchmark.
The company also reported a score of 87.30 on the global olmOCR-Bench, a benchmark used to evaluate document parsing capabilities.
The results are intended to demonstrate the model’s ability to process both Indian-language documents and more general document formats.
Sarvam AI’s emphasis on Indic OCR is particularly relevant to businesses and public-sector organisations handling documents in multiple Indian languages.
Training combines synthetic and real-world documents
Sarvam AI said Vision 2.1 was trained using a combination of synthetic and real-world document scans.
The training process placed particular emphasis on regional handwriting variations and form-extraction scenarios. The model was subsequently fine-tuned using supervised learning.
The company then applied reinforcement learning with verifiable rewards (RLVR), a technique designed to encourage the model to produce outputs that are structurally and textually accurate.
For document-processing applications, this distinction is important because an output that merely appears plausible may still contain errors in numbers, fields, table relationships or document structure.
The training approach therefore aims to improve not only the model’s ability to read a document but also its ability to reproduce the information in a usable and consistent format.
Designed for production workflows
Sarvam AI is positioning Vision 2.1 as a document intelligence model rather than simply an OCR system.
The objective is to allow organisations to take information from unstructured scans and convert it into structured data that can be incorporated into existing production workflows.
For businesses dealing with large volumes of paperwork, this can include extracting information from forms, processing financial documents, interpreting tables and digitising documents written in Indian languages.
The model’s ability to handle multi-page documents is also intended to make it useful for workflows where information is distributed across several pages rather than contained in a single image.
Available through Sarvam’s document intelligence APIs
Sarvam Vision 2.1 is available through Sarvam AI’s document intelligence APIs.
The developer tools are designed to help organisations convert multi-page documents into structured text and automatically extract information from forms and tables.
This API-based approach allows developers to integrate document-processing capabilities into existing applications and workflows rather than requiring users to process individual documents manually.
With Vision 2.1, Sarvam AI is seeking to address a broader challenge in enterprise AI: converting the large amount of information stored in unstructured documents into data that can be searched, analysed and used by software systems.
The focus on Indic languages and handwriting also reflects the diverse nature of documents used across India, where digitisation often involves material that goes beyond standard English-language typed documents.