E-Invoice Extractor 0.2.0 — model card

Developer preview for ERP/AP ingestion in your own SageMaker account. Application code is Apache-2.0; recognition models and third-party runtime components retain their individual licenses. See sources and license boundaries.

Scope and output

CFDI 3.3/4.0 I/E and legacy NF-e 4.00 model 55 XML preserve declared decimal strings, party identifiers, selected document fields, rows and reviewed tax paths. Provider-issued national NFS-e 1.01 XML with embedded DPS has a narrower documented mapping. PDF types are CFDI_PRINTED and DANFE_PRINTED; printed appearance does not establish an underlying XML version.

The PDF path transcribes supported labels and single-line product rows. Text-only pages use PDFium geometry; image pages, including hidden text overlays, are rendered for local Tesseract OCR. A bounded PP-OCRv6/NCNN crop recognizer independently verifies selected fields. No training or fine-tuning was performed for this release. Model weights are hash-pinned. Diagnostic recognition confidence is not a calibrated probability of fiscal correctness.

Missing or conflicting values remain null and unsupported inputs return fixed error codes. Source money precision remains a string; nothing balances totals or infers a currency. NF-e XML has no declared currency field, so it remains null. Every PDF returns a review requirement, including complete-looking outputs.

Finite synthetic evidence

The original corpus contains 2,000 fictional raster PDFs in 96 dependent font/size/layout/blur/country families. It is not 2,000 independent customer invoices or a human-reviewed population sample. Four hundred cases became exposed development replay while fixing RFC clipping. The fixed runtime was then evaluated on the 1,600 previously unused cases, with roles and literal printed-name targets declared before their first inference.

Both countries separately pass these objectives. Only 1389/1600 documents match every target. Issuer RFC is 697/800, including 92 nulls and 11 incorrect non-null values; UUID is 755/800 with 45 nulls. Three monetary targets differ and three expected rows do not match exactly. These errors are retained. Separate synthetic XML regression matches 349/349 target fields.

The vendor-patched platform was replayed on all 2,000 now-exposed PDFs: every literal target remained unchanged. Small recognition-diagnostic confidence differences are recorded separately; none crossed its acceptance threshold. This is compatibility evidence, not a new blind accuracy estimate or a general customer-accuracy guarantee. Evaluate permitted invoices from your actual workload before use.

Performance and runtime

On 8 November 2026, the exact release image completed 84/84 serial requests on one network-isolated ml.c5.2xlarge. AWS-region client wall-time p95 was 10.71 ms for XML, 375.38 ms for text PDFs, 3342.17 ms for raster PDFs and 3095.48 ms for degraded raster PDFs. First requests were recorded separately. The finite workload contains 16 original synthetic XML files and 18 one-page PDF fixtures, repeated for diagnostics. At concurrency two, PDF excess was rejected with the documented pdf_busy response; rejected work is not counted as successful throughput.

One PDF worker is admitted at a time. Concurrent excess returns pdf_busy; callers need bounded backoff or a queue. Batch Transform uses SingleRecord and one concurrent transform. Results are finite repeated synthetic measurements, not maximum capacity or an SLA.

The runtime has no outbound network calls or request-time downloads. Native PDF/OCR work runs in a disposable process group with memory, CPU, time, page, geometry, stream and concurrency limits. Source documents are not logged. Responses contain supplied financial data and bounded evidence, so secure your own output storage.

Unsupported: fiscal/signature/issuer validation, complete official-schema conformity, universal layouts, multiline/wrapped/merged rows, handwriting, reliable arbitrary rotation, municipal/other NFS-e variants and new tax-reform semantics. The national NFS-e reference-XSD series regex conflicts with its documented example under libxml2; full XSD conformity is not claimed.

Price and support

USD 0.50 per software instance-hour on ml.c5.2xlarge; 14-day software trial for eligible buyers. AWS infrastructure and storage remain separately billable during the trial. Support: info@softwaresushi.com; provide redacted structure and fixed error codes, not taxpayer data.