Document extraction

Open /extraction in the Next.js playground. Paste document text, select fields, and choose Run extraction. The included invoice has an issue date, due date, vendor, subtotal, tax, and final amount so the ranking is visible.

The prototype uses three functions in src/extraction/extraction.ts:

• extractCandidates(text) finds source spans with regular expressions. It returns candidates for date, counterparty, amount, and document_type, including exact values, character offsets, and short source snippets.
• rankWithJev(field, candidates, text) submits a choice question containing candidate IDs and an explicit null choice. It maps the selected ID back to the source value. An unknown ID or invalid probability is an error.
• runExtraction(text, { fields?, signal? }) extracts once and ranks up to three fields in parallel. Failures remain visible per field; completed fields survive.

Jev does not generate text. Field instructions favor the issue date, issuer or vendor, and final amount; they prohibit inference, calculation, and following instructions embedded in the document. Evidence is copied locally, not generated. The TypeSafe choice API returns probabilities over named criteria. The UI displays the selected choice's probability and, when supplied, Jev's separate confidence. Neither is a guarantee that the field is correct.

If no candidates exist, the field returns null locally without a network call or invented confidence. A model-selected null has its own returned probability. Request errors are shown as Failed, not as a successful null extraction.

Run locally

Requires Node.js 22 or newer:

pnpm install --frozen-lockfile
cp .env.example .env.local
# Edit .env.local and set TYPESAFE_API_KEY to your key.
pnpm dev

Open http://localhost:3042/extraction (or the port printed by Next.js). The server reads TYPESAFE_API_KEY; it is never sent to the browser. .env.local is ignored by Git. Do not use a NEXT_PUBLIC_ variable for this key.

pnpm test
pnpm typecheck
pnpm build

The shared /api/run Next.js route validates requests, bounds request/response sizes, attaches authorization, and calls TypeSafe. Text is sent to TypeSafe only when you run the extraction.

Scope

The rules handle common English dates, labeled vendors or company-name lines, USD amounts, and explicitly printed document types. They are a prototype, not OCR or a universal invoice parser. Values retain their source formatting. Identical values are deduplicated and use their first source span. Documents are limited to 40,000 characters and each field to 100 distinct candidates. Oversized documents are rejected explicitly; split them into smaller sections. No candidates are silently dropped. Ambiguous documents may return null.