Tax Doc Classifier — Sathvik2005

Install Node 20+ and pnpm, plus poppler for pdftotext and pdfinfo. Then run pnpm install, pnpm typecheck and pnpm test, export TYPESAFE_API_KEY, and call classifyPage with jevBackend() and the criteria JSON. It can also be added to another project as a dependency.

Yes. jevBackend() reads a TypeSafe API key from the environment or a git-ignored .env file and never writes it to disk or logs it. Reported cost is about $0.00115 per page, versus $0.039 per page for the Sonnet classifier.

The page's text, split into header, body and footer, is sent to Jev, which returns a probability over options instead of text. PDF parsing happens locally through pdftotext, and the form criteria file is generated locally from IRS PDFs.

On 314 TaxCalcBench filled pages it reported 0 wrong and 0 strict errors at $0.36. On 753 blank IRS forms covering 261 forms it reported 38 strict errors (5.05%), all low confidence rather than wrong answers.

A page is counted as an error if the answer is wrong or its confidence falls below 0.95. Each run prints the strict score, lists every strict error with its confidence and runner-up, and writes per-page results to eval/results/.

Only classifyPage on lines of text works without poppler. The PDF helpers pdfPageLines and the corpus download rely on pdftotext and pdfinfo, which come from poppler-utils on Debian or Ubuntu and from brew install poppler on macOS.

One request picks among 7 page kinds and a 230-option form list. Five corporate or foreign forms that absorb schedules get a second smaller question. formConfidence is the minimum over steps: above the gate you act, below it you fall back.