Unsiloed AI provides a document extraction service that returns not only the recognized text but also two confidence scores for each field: an extraction_score measuring how well the value was read and a grounding_score indicating how confident the model is that the value was located in the document. These scores help identify readings that may be affected by OCR issues such as smudges or low contrast. The service also supplies a citation pointing to the exact page and bounding box of the source text, allowing reviewers to trace each extracted value back to the original document.

TypeSafe’s Jev complements this extraction by turning focused questions into typed answers. In the document‑review workflow, each extracted field is sent to Jev along with a question such as 'Is the extracted dose and frequency plausible on its face for this medicine and strength?' Jev’s Noul question type returns a probability between 0 and 1, where values near 0 or 1 indicate a clear yes or no and values near 0.5 signal uncertainty. The model does not see the underlying image; it judges solely on the extracted text and any supplied context, making its assessment independent of OCR quality.

Diagram showing Jev question contract with Noul, Choice, and Score question types and their outputs
Illustration of how Jev receives focused questions and returns Noul probability, Choice distribution, and Score confidence for document‑review validation.www.unsiloed.ai

This image depicts the Jev API contract described in the article, where Noul returns a yes‑no probability used to flag implausible extracted fields.

The review pipeline combines the two signals: a row is flagged for review if either the lowest Unsiloed score across its fields falls below 0.90 or the Jev Noul answer drops below 0.50. This design catches errors that either model alone might miss—for example, a coffee‑stained cetirizine dose that Unsiloed extracts as “1 tablet” with low confidence (extraction_score 0.74, grounding_score 0.76) but Jev still deems plausible (noul 0.94), or the clearly printed ibuprofen instruction that reads “10 tablets, three times daily” with high Unsiloed scores (0.99) but Jev rejects as implausible (noul 0.04).

Screenshot of a medication list with a coffee stain obscuring the cetirizine dose and the ibuprofen instruction showing 10 tablets three times daily
Example prescription document used in the article to illustrate OCR‑affected and clearly printed but incorrect instructions.www.unsiloed.ai

The image shows the coffee‑stained cetirizine dose and the ibuprofen instruction cited as examples of errors caught by the Unsiloed and Jev pipeline.