varga

Two questions about one document. guess_type answers the first from cue phrases and countable evidence, with no model unless two kinds tie. as_schema and structured answer the second against a shape you name or invent.

Install

uv add varga            # doctypes; no model, no network
uv add varga rishi      # ...and a chat for `structured` to ask

There are no extras: structured imports what it needs when it is called.

What is it

INVOICE = '''# INVOICE
Invoice No: ACM-2024-0117
Date: 2024-03-01
Bill to: Contoso GmbH, Berlin
From: Acme Supplies Ltd
Total due: 1,240.00 EUR
Payment terms: net 30
'''
r = guess_type(INVOICE)
r.doctype, r.decisive, round(r.margin, 2)

('invoice', True, 0.4)

decisive is the whole point: it says whether the cues settled it. When two kinds tie, margin is 0 and a caller knows to spend a model call. Nothing else here costs one.

signals(INVOICE).counts

{'money': 1, 'date': 1, 'ref': 1, 'heading': 1, 'org': 2}

What is in it

from dataclasses import fields
[f.name for f in fields(as_schema('invoice'))]

['number',
 'date',
 'due_date',
 'vendor',
 'vendor_tax_id',
 'bill_to',
 'ship_to',
 'currency',
 'subtotal',
 'tax',
 'total',
 'payment_terms',
 'items']

Fifteen shapes are built in: invoice, purchase_order, quote, receipt, catalogue, contract, resume, paper, meeting_notes, id_card, passport, bank_statement, payslip, tax_return, other. Or write one on the spot:

[(f.name, f.type.__name__) for f in fields(as_schema('vendor:str, total:float, items:list'))]

[('vendor', 'str'), ('total', 'float'), ('items', 'list')]

structured(chat, prompt, schema) fills it. It prefers the model’s constrained mode and falls back to a JSON reply when that raises, so a model without tool calling still answers.

structured(..., cite=True) returns {field: {value, quote}}, the quote copied from the document, and locate(fields, pages) turns each quote into the page and block it came from with a string match, not a model.

The doctypes

len(DOCTYPES), sorted(DOCTYPES)

Ten of them are work-product labels: proposal, presentation, requirements_spec, technical_design, regulatory_guidance, procedure, qa_artifact, roadmap, claim, clinical_record, and each is decisive on its own text without a model. Prose that merely borrows the vocabulary scores under 0.2 against every one of them.

Six are sensitive: id_card, passport, bank_statement, payslip, tax_return and clinical_record. They are in SENSITIVE, and guess_type(...).sensitive says when the answer is one of them, so a caller can keep that document on the machine.

What it depends on

fastcore and rahasya, for the honorific-anchored name regex the entity leg shares with the privacy gate. One definition of what a name is.

rishi is needed only by structured, for the chat it is handed and the JSON fallback.

Development

The notebooks in nbs/ are the source; the modules are generated.

pip install -e .
nbdev_prepare