多模态、OCR 与文档智能测试教程

// 金标应有 100 个金额字段；系统输出 98 个，其中 95 个值与证据都正确。
precision = 95 / 98;          // 96.94%
recall = 95 / 100;            // 95.00%
f1 = 2 * precision * recall / (precision + recall); // 95.96%

// 仍要单列 P0 错误：任何一张发票把 1,000.00 识别成 10,000.00 都可硬失败。

const moneyOk = Math.abs(netAmount + taxAmount - grossAmount) <= 0.01;
const detailOk = Math.abs(
  lines.reduce((sum, line) => sum + line.quantity * line.unitPrice, 0) - grossAmount
) <= 0.01;
const dateOk = issuedAt <= reimbursedAt && accountingPeriod.includes(issuedAt);

route = moneyOk && detailOk && dateOk ? "AUTO_PASS" : "HUMAN_REVIEW";

given({ pdfRenderer: "drop_page_2", ocr: "ok", database: "ok" });
const run = await processBatch("BATCH-0810");

expect(run.files[0].status).toBe("BLOCKED");
expect(run.files[0].errorCode).toBe("PAGE_COUNT_MISMATCH");
expect(run.databaseRowsWritten).toBe(0);
expect(run.audit.originalFileHash).toBeDefined();

{
  "batch_id": "BATCH-0810", "source_files": 20, "source_pages": 57,
  "accepted": 14, "human_review": 5, "rejected": 1,
  "reviewed_business_keys": 19, "inserted": 18, "idempotent_hits": 1,
  "approved_amount": "126430.55", "inserted_amount": "126430.55",
  "currency": "CNY", "difference": "0.00",
  "model_version": "ocr-v7", "schema_version": "invoice-v4",
  "reconciliation_status": "PASS"
}