Dharma targets structured fields beyond OCR

Typed output moves document extraction past OCR: field names, types and nested objects become part of the application's input contract. In an article dated November 4, 2026, Dharma AI introduces its field extractor, which takes document images and returns those structures.

The distinction from basic OCR is semantic organization. The system can work against a supplied schema or extract available fields. For a document containing line items, the useful result associates descriptions and values correctly rather than returning a stream of text whose relationships the application must reconstruct.

That can remove downstream work, but it moves the quality question. Valid JSON may still assign a value to the wrong field, confuse an item relationship or turn missing information into an unwarranted answer. Structural validity and factual correctness need separate checks.

Dharma's reported evaluation belongs to the team's own setup. For an integration, compare representative outputs with the images, including absent fields and ambiguous layouts; these are proposed checks, not results obtained here. The parser can establish that an object is readable. The remaining implementation question is whether its field assignments preserve the document's meaning for the next business step.