tax-doc-classifier

A tax document classifier built with Jev.

We ingest thousands of tax documents using an LLM pipeline built last tax season. Jev classifies 100% of our tax document corpus at $0.001 per page — 34× cheaper and 6× faster than that LLM setup. This is the classifier, open sourced.

One request per page. The page's text goes to Jev, TypeSafe's decision model, which returns a probability over 261 IRS forms and 7 page kinds instead of text. No model is trained and nothing is hosted: the classifier is a JSON file describing each form, generated from the IRS's own PDFs.

import { classifyPage, jevBackend, pdfPageLines } from 'tax-doc-classifier'
import criteria from 'tax-doc-classifier/data/criteria.json' with { type: 'json' }

const lines = await pdfPageLines('return.pdf', 3)
const r = await classifyPage(lines, { backend: jevBackend(), criteria })
// r.form            'form-1040-schedule-a'
// r.kind            'form_page'
// r.formConfidence  0.99
// r.gated           true   (formConfidence >= 0.95)

No model is trained and nothing is hosted. The classifier is a JSON file of form descriptions (data/criteria.json) and a decision model that reads a page and picks from them. The decision model is TypeSafe Jev, a System One model that returns a probability over options instead of text.

Setup

1. Node 20+ and pnpm.
2. poppler — provides pdftotext and pdfinfo, used to read PDF pages. Only classifyPage on lines of text works without it.
   • macOS: brew install poppler
   • Debian/Ubuntu: apt install poppler-utils
3. A TypeSafe API key. Get one at typesafe.ai, then:

   export TYPESAFE_API_KEY=…
   

   or put it in a .env file (git-ignored) and load it with your usual tooling. The key is read from the environment by jevBackend(); it is never written to disk or logged.
4. Install and check:

   pnpm install
   pnpm typecheck
   pnpm test
   

To use it as a dependency in another project:

pnpm add github:<org>/tax-doc-classifier

The published package ships dist/ and data/criteria.json; pdftotext is still required on the host for the PDF helpers.

Results

The score is strict: a page counts as an error if the answer is wrong or its confidence is below 0.95. Every number below is reproducible with pnpm eval.

Corpus	Pages	Forms	Wrong	Strict errors	Cost
TaxCalcBench filled forms (W-2, 1040, 1099-x, 1098-x)	314	15	0	0 (0.00%)	$0.36
Blank IRS forms (every TY2025 MeF 1040-series form + 40 information returns)	753	261	0	38 (5.05%)	$0.86

The 38 low-confidence pages on the blank corpus are instruction pages that name no form, deep pages of corporate forms (5471, 8865, 1118), and single-schedule forms hedging against their parent. None is wrong.

Against the LLM setup it replaces

Measured on the same bench pages, same machine, same hour. The previous classifier sends each page as a PDF to Claude Sonnet with the form registry in the prompt; different pages never share the prompt cache, so every page pays the full prompt.

Per page	Sonnet classifier	Jev classifier
Cost	$0.039	$0.00115	34× cheaper
Latency, warm	~3.3 s	~0.5 s	6× faster
Forms it can name	30 (phrase table)	261

How it works

PDF page ──pdftotext──▶ { header, body, footer }
                              │
                              ▼  one request
                     kind  : 7 options
                     form  : 230 options + not_in_this_list
                              │
                              ▼  second request only if form ∈ {5471, 8865, 8933, 1118, 5713}
                     sub   : that form and its schedules
                              │
                              ▼
        { form, kind, formConfidence, gated, probabilities }

• A blank page is answered without a call.
• The 230 first-list options are every form as the page prints it. Five corporate/foreign forms absorb their 35 schedules and get a second, small question; the hierarchy exists only where the page announces the parent more clearly than the schedule.
• formConfidence is the minimum over steps. Gate on it: above the gate, act; below, fall back to whatever you use today.

Page kinds

form_page · instructions · blank · cover_sheet · state_tax_form · broker_or_bank_statement · letter_or_other

The last three route a page away from the federal form list; they are defined but not yet evaluated.

Form ids

Ids follow the IRS MeF naming, kebab-cased, parent first:

form-1040                 form-1040-schedule-a        form-1040-nr-schedule-nec
form-1099-int             form-w-2g                   form-5471-schedule-j
form-1065-schedule-k-1    form-8995-a-schedule-a

Form 1040 Schedule A in the IRS accepted-forms list becomes form-1040-schedule-a, mechanically and in both directions (mefNameToId). Every id matches ID_GRAMMAR, and the test suite refuses a criteria.json that contains one that does not.

A page number, when a caller asks for it, is a suffix: form-1040/p2. Every prefix of an id is a valid, less specific id. No suffix means the page was not asked, never "page 1".

The data file

data/criteria.json is generated — never edited by hand:

pnpm build-criteria

reads the IRS accepted forms and schedules XLSX, downloads each form's PDF from irs.gov, and extracts per form: the printed label, the title, the page count, the parent (for schedules), up to 8 box labels (information returns only), and the sibling forms it must not be confused with. Re-run it when the IRS publishes new revisions.

Evaluation

pnpm eval:download                                 # IRS PDFs → eval/corpus/blank
TYPESAFE_API_KEY=… pnpm eval --corpus=blank
TYPESAFE_API_KEY=… pnpm eval --corpus=bench --dir=/path/to/tax_calc_bench/ty25/test_data

Each run prints the strict score, lists every strict error with its confidence and runner-up, and writes the full per-page results to eval/results/.

The bench corpus is TaxCalcBench and is not redistributed here.

Plugging in another model

Backend is one method: ask(state, questions) → { answers, inputTokens } where each answer is a probability distribution over the question's options. jevBackend is the only implementation shipped. Anything that can return calibrated probabilities over a labelled option set can implement it.

Limits

• Text only. Scanned pages need OCR first; a page with no text layer is reported as blank.
• Federal forms only. State forms are detected by kind but not identified.
• English only.
• Requires pdftotext and pdfinfo (poppler) on the path for the PDF helpers; classifyPage itself takes lines of text and needs neither.
• The confidence is calibrated on the corpora above. Calibrate on your own pages before choosing a gate.

Requirements

Node 20+, poppler, a TypeSafe API key in TYPESAFE_API_KEY.

Licence

Apache-2.0. See DATA-LICENSE.md for the IRS-derived data.