An invoice parser needs to preserve a missing field as carefully as it reads a printed total.
Dharma-AI described its Field Extractor in a November 4 technical article. The specialist model takes a document image and returns structured fields. Users can ask for what the model finds or provide a schema with required key names and types.
A schema changes the output contract
The provider describes nested fields, table rows, and null values for requested information absent from a document. That makes the intended output more specific than an OCR transcript. An application still needs to validate the result before inserting it into a business system.
Consider an invoice whose payment terms appear without a due date. A schema may require both fields, but that requirement does not supply the missing date. The application should preserve the difference between text printed on the page and a value calculated from other information.
The benchmark combines several kinds of correctness
Dharma reports a 2-billion-parameter model and a benchmark of 546 documents under three request scenarios, producing 1,638 cases. Its score combines field matching, similarity of the extracted values, and valid JSON output. The scenarios include deliberately requesting non-existent fields.
The headline score therefore combines several possible failure modes. A wrong line-item position, a damaged value, and an invalid response can each affect the outcome. For a particular workflow, separate those errors so a high aggregate score does not hide the one that breaks an import.
The reported saving is not a price quotation
Dharma reports lower cost than the frontier API configurations in its comparison. Its own estimate uses GPU time on an L40S priced at US$2.24 per hour. The API side uses measured token consumption and published list rates. These are provider-reported measurements, not a benchmark we reproduced.
A self-hosted budget must also account for utilization, queueing, deployment work, and monitoring. A GPU waiting between small batches can cost more per page than a continuously busy benchmark. Compare the cost per accepted extraction under the workload you actually expect.
Test absent fields before connecting a database
A small evaluation set could include an invoice missing a due date, a receipt with a faint total, and a multi-row table. Validate required keys, types, row order, and the handling of missing values. Keep the original image alongside corrections for review.
We have not run those tests or established a public checkpoint or deployment license for this extractor. Our Gemini document-processing article covers a different ingestion issue: what evidence a file actually supplies. Which extraction error would cause the most damage in your workflow?
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
