Summary

  • Nutrient announced general availability of its Data Extraction API on September 10, combining schema-defined fields with source-location evidence.
  • Its documentation says confidence is not a probability and a missing score is not low confidence. A grounded value can still require business review.

A number can be found exactly on an invoice and still be the wrong number to pay. It might be a line item rather than the total the buyer requested. That hypothetical distinction—not a reported Nutrient failure—is where document extraction becomes an operating decision.

Nutrient's September 10 announcement offers enterprises an API for parsing documents and extracting structured fields. The useful change for a workflow is not merely that a model produces a value. It can also provide a route back to the page, letting a reviewer inspect the evidence behind that value.

The extract endpoint guide separates that job from full-document parsing. Extraction requires a JSON Schema, runs a parsing stage internally and returns the declared data fields. Citation metadata accompanies the data, with page information. The citation contract adds an important qualification: citations are on by default, but disabling them skips their computation and leaves the metadata empty. An empty citation object therefore does not, by itself, mean the extracted value could not be found.

When citations are present, their match labels distinguish exact single-block matches, multiple-block matches, partial resolution, approximate matches and values that could not be grounded to a source location. The guide suggests routing fuzzy_match and not_found to review. That is guidance for building handling logic, not proof that a ready-made approval process has already cleared a customer's documents.

Confidence deserves separate treatment. It is a relative signal between zero and one, not a calibrated probability or a percentage. If the engine supplies no score, the field is omitted: absence means unavailable, not zero or low confidence. The documentation also calls for human review of high-stakes fields regardless of score. Locating an amount and authorising its use are different acts.

The headline performance number measures something else again. Nutrient's benchmark page reports an overall score of 0.932 for understand mode on 200 PDFs, using reading order, table structure and heading hierarchy. It describes understand and agentic evaluations as internal runs on a public corpus; public listing applies to its text/structure open-source tool. This is not a 93.2% field-correctness guarantee. Its local processing timings are not API roundtrip latency, and the fastest mode's speed cannot simply be attached to another mode's accuracy.

Even the product showcase draws a useful line: recording an evidence audit does not mean approving the result. Its decision-gate videos illustrate a workflow concept, not shipped UI. Traceability can make a mistake easier to find; whether it saves work depends on what happens after it is found.