Document Review with Unsiloed and TypeSafe Jev

The document review process using Unsiloed and TypeSafe's Jev system begins with an agent storefront that provides necessary information for setting up agents to purchase autonomously or with approval. The storefront handles payments without requiring direct API calls. An extraction pipeline can extract structured fields from documents, but it may still pass bad data downstream. Using Unsiloed and TypeSafe's Jev together helps to identify OCR failures and general mistakes in the document.

Unsiloed extracts fields and locates them on the page, while Jev judges the extracted information against focused questions. The extracted information, state, and context needed to interpret them are contained in the state. Each question tells Jev what to check and which answer type to return. The answers can be Noul, Choice, or Score, each with a probability distribution and confidence value.

In a medication list example with six rows, one dose is obscured by a coffee stain, while another row contains implausible quantities. Unsiloed preserves what is printed and provides confidence scores to identify ambiguous readings. A Noul value of 0.04 indicates a 4% probability of a statement being true, not confidenc e in Jev's decisions. Values near either end indicate a clear direction, while values near the middle deserve caution.

After extraction, Jev's state is constructed using the extracted values and citations, which identify the page and bounding box of the source text. Each field has two Unsiloed confidence scores: extraction_score and grounding_score, both ranging from zero to one. For each medication row, a Noul question is asked: is this dose and frequency plausible for this medicine? Jev evaluates the instructions independently for each row.

In the server-side module, a JSON request is constructed and posted to the TypeSafe API, authenticated with the TypeSafe API key. The answer is then matched with the saved extraction using the shared ID. The code decides whether the row needs review by taking the lowest Unsiloed score across the row's fields and using Jev's noul value for plausibility checks.

These demo rules flag rows with missing or invalid scores. The original extraction and Jev request and response are saved alongside the reviewer's decision for future reference.