The Signal · Issue 01 · India & Emerging Markets
Reading India's paperwork, in 22 languages
Sarvam's Vision 2.1 targets the last mile of document intelligence: Indic handwriting, forms and multi-page tables. Its most useful contribution may be the open benchmark that came with it.
Facts are attributed and linked to their sources. The analysis and opinions are the author's own.
What it is
On 24 September Bengaluru-based Sarvam AI released Sarvam Vision 2.1, an update to the document-intelligence model it launched in February 2026 as part of its sovereign models line-up. The release adds structured key-value extraction from forms and tables and Indic handwritten recognition, and Sarvam says it addresses earlier weaknesses around hallucinations and inconsistencies.
Alongside the model, Sarvam released an Indic OCR benchmark on Hugging Face with 6,909 samples, 6,609 spanning the 22 official Indian languages and 300 in English, drawn from sources such as newspapers, brochures, textbooks and historical writing from 1800 to the present. On that benchmark Sarvam reports 87.39% overall accuracy for Vision 2.1, ahead of the other systems it evaluated.
Sarvam frames document intelligence as a multi-objective problem and reports a Pareto analysis rather than a single score, noting that a model strong on English benchmarks often performs poorly on Indic languages. It also observes that the global benchmarks it reports are arguably saturated.
Why it matters
Much of India's financial and public life still runs on documents: KYC forms, invoices, land records, handwritten applications. Errors at the extraction step flow silently into every model downstream. A model that reads Indic handwriting well is not a curiosity; it is a data-quality control for half the country's paperwork.
Errors at the extraction step flow silently into every model downstream. Better OCR is a data-quality control.
The benchmark matters as much as the model. A public, language-by-language test set lets others check claims and track progress, and Sarvam's own note that a vendor-reported benchmark should be read alongside real-world testing is the right caution.
The Pareto framing is a good habit for any team evaluating document systems. A single accuracy score hides trade-offs: one model is better on old scans, another on tables, another on a particular script. The right question is which system is not beaten everywhere by another, and then which trade-off fits your documents. That is how we should buy models in general.
For lenders and insurers the use cases are obvious and the controls are not optional. Extracted fields from an identity document or an income proof feed eligibility and pricing decisions. I would want field-level confidence scores, a human review queue for low-confidence fields, and a regular sample audit by language, because the reported accuracy varies a lot between scripts.
The broader point is about who builds evaluation infrastructure for Indian languages. A public benchmark with thousands of samples across 22 languages is a public good, whoever builds it, and it raises the bar for every vendor serving this market.
What I'd watch
- Independent results on the Indic benchmark from teams other than the vendor.
- Per-language error rates in production. The reported scores vary widely by language; low-resource scripts need their own thresholds.
- How extracted personal data from forms is handled under DPDP obligations once these pipelines scale.
Sources
- Sarvam Vision 2.1: Pushing the Pareto frontier of document intelligence. Sarvam AI, 24 Sep 2026.