Rule document-extractor

Apply this rule whenever work touches:

• libs/shared/document-extractor/**/*.ts
• libs/shared/text-extractor/**/*.ts

The document extractor framework separates concerns between parsing (extracting fields from documents) and evaluation (deciding which fields matter). Parsers are deliberately field-neutral.

Parser design

Parsers implement DocumentParser<T> and extract as many fields as possible without enforcing which ones are required:

import type { DocumentParser, NonEmptyString } from '@carrot-fndn/shared/document-extractor';

export class InvoiceParser implements DocumentParser<InvoiceFields> {
  parse(rawText: string, layoutId: NonEmptyString): Partial<InvoiceFields> {
    return {
      invoiceNumber: this.extractInvoiceNumber(rawText),  // string | undefined
      issueDate: this.extractDate(rawText),               // string | undefined
      totalAmount: this.extractAmount(rawText),            // number | undefined
    };
  }
}

Every field in the return type is optional. If extraction fails for a field, return undefined.

Review required

The reviewRequired flag signals that a human should verify the extraction. It is triggered exclusively by extraction quality signals:

• A field was extracted with low confidence
• The layout match score is below 0.5

// Correct - confidence-based review trigger
const reviewRequired = matchScore < 0.5 || fields.some(f => f.confidence < THRESHOLD);

// Wrong - business rule in parser
const reviewRequired = !fields.totalAmount || fields.totalAmount < 100;

Layout identifiers

Layout IDs and layout names use the NonEmptyString type. Cast string literals explicitly:

import type { NonEmptyString } from '@carrot-fndn/shared/document-extractor';

const LAYOUT_ID = 'invoice-standard-v2' as NonEmptyString;

Default layouts

Only register default layouts in defaults.ts for document types that are stable and well-established. Experimental or rapidly evolving formats should be configured at the processor level instead.