Document Review With Unsiloed and TypeSafe Jev

Description and use

Combining Unsiloed and TypeSafe Jev enables detection of OCR errors and plausibility checks of medication doses, routing only questionable rows for manual review.

Problem

An extraction pipeline can read a value correctly and still pass bad data downstream; OCR failures can leave incorrect or ambiguous data in extracted fields.

Outcome

Combining Unsiloed extraction with Jev plausibility checks allows targeted review of only the rows with low confidence or implausible data, improving efficiency and accuracy of medication reconciliation.

Workflow

• 1. Extract medication list with Unsiloed using a schema that returns dose, frequency, route, and instructions with confidence scores and citations.
• 2. Assign each row an ID and retain citations for traceability.
• 3. Construct a Jev request with state containing context, careContext, and medications array; include one Noul question per row asking if the dose and frequency are plausible.
• 4. Send the request to Jev (model jev-1.13.0).
• 5. Receive Noul answers (probabilities) and route rows with low plausibility to human review.
• 6. Review original document using citations to verify or correct extraction.

Evidence

An extraction pipeline can read a value correctly and still pass bad data downstream. A clearly printed extra zero survives excellent optical character recognition (OCR).

Source

Using Unsiloed and TypeSafe’s Jev together helps us check for OCR failures and general mistakes in a document.

Source

We ask one Noul question per row: is this dose and frequency plausible for this medicine? All six rows share one request, with each question evaluated independently.

Source