Sarvam Vision 2.1 Adds Structured Form Extraction and Handwriting Recognition for Indian Languages

Sarvam AI announced Vision 2.1 on September 24, adding cross-page table processing, form key-value extraction, and handwriting recognition for English and 22 Indian languages. The update extends document recognition to field extraction, making it worth evaluating for tasks that could reduce manual form-processing work. [Official announcement](https://www.sarvam.ai/blogs/sarvam-vision-2-1)

The system retains its design of pairing a vision-language model with specialized processing components, using a semantic layout parser and a pointer-based reading-order network to assist recognition. Training combines real and synthetic data, followed by supervised fine-tuning and reinforcement learning with verifiable rewards. The reported performance should therefore be understood as the result of the entire document-processing system, rather than attributed directly to a single model. [Architecture overview](https://www.sarvam.ai/blogs/sarvam-vision-2-1)

The integration specifications are now concrete: the Extract API accepts a JSON Schema or a saved extraction configuration, runs asynchronously, and returns field results with confidence values and source information for each leaf field. Schemas can be nested up to four levels, and each field must have a type and description. This can help engineering teams plan human review, though they should separately verify whether the confidence values accurately reflect error probabilities. [Extraction API documentation](https://docs.sarvam.ai/api-reference/doc-ai/job/extract)

For teams building search or retrieval-augmented generation datasets, the Digitise API can preserve layout and output HTML, Markdown, or JSON containing bounding boxes for text blocks. This interface allows teams to retain the relationship between text and its location in the original document, making it easier to trace data sources. Retrieval quality still needs to be evaluated independently. [Digitise API documentation](https://docs.sarvam.ai/api-reference/doc-ai/job/digitise)

Sarvam reports an overall score of 87.3 on olmOCR-Bench. This release uses the full official test set, while the previous version used a filtered English subset, so the two scores cannot be directly subtracted to measure improvement. This is also a vendor-run evaluation, and its results do not establish how well the system handles Chinese documents. [Evaluation scope](https://www.sarvam.ai/blogs/sarvam-vision-2-1)

Another reproducible resource is the public Indic OCR Bench: it contains 6,909 text blocks across English and 22 Indian languages, with reference answers reviewed twice by language experts. Scoring converts word error rate into word accuracy after normalizing the text; the “valid sample” metric excludes uncontrolled repeated output. Reproductions should report the overall scores, failure counts, and results for each language. Block-level recognition scores do not mean that every field in a complete form is correct. [Dataset and scoring details](https://huggingface.co/datasets/sarvamai/indic-ocr-bench)

There is also a version-mapping issue to resolve before deployment: the current documentation still lists `sarvam-vision-v1` as the model default, which is not enough to confirm that requests use version 2.1. Developers should verify the server-side version and test field accuracy on their own cross-page tables, partial-failure handling, and end-to-end latency before deciding whether to integrate the API into production workflows. [API model parameters](https://docs.sarvam.ai/api-reference/doc-ai/job/extract)