Dharma AI described its field extractor on November 4.

The big change

  • Schemas specify fields; open requests do not.
  • Dharma reports higher benchmark scores.
  • A trial is linked; weights are unconfirmed.

From a page to fields

Dharma says invoice schemas can request supplier, number, due date and total. Open requests identify fields instead. Both return JSON; missing requested fields should return null.

A schema is useful because a receiving application needs predictable names and types. The JSON Schema project explains that a validator can check whether data conforms to declared structural rules. Valid structure alone does not establish that an extracted value matches the page; users still need a way to assess accuracy.

What was measured

Dharma tested 546 documents in three scenarios—open, existing fields and absent fields—making 1,638 cases. Its composite combines field matching, value similarity and valid JSON. It reports 0.834 against Claude Opus 5’s 0.816, with low or adaptive-low comparator effort; this is not independent verification.

Dharma reports costs 6–147 times lower: provider list prices applied to measured tokens versus self-hosted GPU time on one L40S at $2.24 per hour. Hardware, utilization and latency affect costs; this excludes broader deployment costs and is not a product price.

Dharma says a LoRA trained on extraction examples adapts its two-billion-parameter OCR-LITE v1.1 checkpoint. The older OCR model has listed weights; its paper does not test Field Extractor. BIG CHANGE did not test it.