Document Triage: Type and Urgency on Two Axes

#The scenario

A shared intake folder or inbox receives every kind of document the business touches: contracts awaiting signature, invoices awaiting payment, resumes awaiting review, support attachments awaiting response. Each type has a different owner, a different pipeline, and a different clock — an invoice has a payment deadline, a contract may have a signing window, a resume has a candidate going cold. A human sorting this pile is slow; a single "importance" score conflates what it is with how fast it needs handling.

#Why Jev fits

Document triage is two orthogonal closed-world judgments, and Jev lets you ask both in one call: a Choice over document types, and a Score for urgency. Two typed questions, one request, structured output you can route on directly — no parsing free text, no prompt glue. At a reported $0.042 per million input tokens with output free (early-access pricing — verify at jev.com), triaging every document at arrival is a rounding error next to the cost of one misrouted contract.

#Judgment design example

Structured input per document:

Judgments to define:

1. Choice (type): "What kind of document is this?" → contract / invoice / resume / support_ticket / other. Type determines the destination pipeline: extraction, approval flow, or human queue.
2. Score (urgency): "How time-sensitive is this, 1–5?" → a renewal contract expiring this week outranks a speculative resume, regardless of arrival order.
3. Noul (safety): "Does this document appear to be corrupted, blank, or mis-scanned?" → yes → reject to a repair queue before any pipeline wastes effort on it.

#Where the LLM sits

Downstream, per pipeline, on the survivors. Contracts route to a clause-extraction LLM, invoices to a line-item extractor, resumes to a screening summary — each with a prompt specialized for that document type instead of one universal "handle anything" prompt. Jev's verdict decides which prompt and which model tier each document deserves; urgent + complex can justify a frontier model, routine can use a small one.

#Watch-outs

• Judge on extracted text, not the binary. Jev currently accepts text (and structured data), not images or PDFs — a scanned document needs OCR first, and OCR garbage will poison the verdict. That's what the corrupted-document Noul is for.
• Type and urgency are independent axes. Don't collapse them into one "priority" question — a low-urgency contract and a high-urgency invoice need different pipelines, and you lose that information if you only ask "how important."
• Long documents need excerpts. Context is reported at 64k tokens with the longest single question capped at 32k; a 200-page contract gets first-page + signature-block excerpts, not a full dump.

#Sources

• jev101.com — application-scenario coverage of document/mail triage pipelines (邮件分流、工单处理) in its use-case library.
• learnjev.com — Three Primitives tutorial (mixing Choice + Score + Noul in one request) and System One concepts.
• jevai.wiki — API reference: typed questions map, context limits, and model pages.
• "Jev: The New AI Model That's Breaking The Internet (Full Tutorial)" — Moritz (~43K views) — structured input → typed verdict request shape.

Unofficial fan-made handbook. Not affiliated with TypeSafe AI or jev.com.