Document Review with Unsiloed and TypeSafe Jev

This describes a practical workflow that pairs Unsiloed’s document extraction with TypeSafe’s Jev to catch OCR and plausibility errors in structured data. Unsiloed extracts fields and returns extraction_score and grounding_score plus page and bounding-box citations; Jev evaluates those extracted values against focused, typed questions. Jev separates state (the extracted records and context) from questions (which specify answer types and criteria). Three answer types exist - noul (P(yes)), choice, and score - each returning probabilities and, for choice/score, a concentration-based confidence. The example uses noul questions per medication row to judge whether dose and frequency are plausible in the stated care context.

The implementation sends all rows and context to Jev, matches answers to row IDs, and combines Jev’s plausibility (noul) with Unsiloed’s lowest field score per row to decide review routing. Example rules flag rows with reading < 0.90 or plausibility < 0.50; a stained cetirizine read as “1 tablet” had low extraction scores but high noul, so human review verifies the obscured glyph, while an ibuprofen row printed “10 tablets three times daily” had high extraction scores but low noul and also requires review. All requests, responses, citations, and reviewer resolutions are stored for audit. Estimated Jev input-token cost per sample is negligible (~$0.000076 per document at published rates).