Document API: extraction with field evidence
Review what can be extracted from a document, where it comes from and what requires a human decision.
From your own file to a traceable decision
1. Upload a bounded document
Choose a text-based PDF or UTF-8 CFDI 4 XML. Only a PDF accepts an optional comparison XML. The dashboard validates format, size and access; originals remain private within your account and project.
2. Inspect fields and provenance
Inspect candidates, source page text or XPath and attribute, crosschecks and arithmetic. Ambiguous or unextracted fields remain explicit; no text, coordinates or values are invented.
3. Review all eleven fields
Accept a cited candidate, mark a field unknown or record a human correction with a note. Version control detects concurrent reviews and preserves the original extraction.
Current service configuration
Capabilities reflect the published catalog; they do not grant account access. Operation requires a pilot grant or current paid project period, as well as available processing and isolated parsing; a configuration flag does not guarantee immediate capacity.
Text-based PDF extraction
Checking…CFDI 4 XML attribute extraction
Checking…Field provenance and candidates
Checking…Crosschecks and arithmetic consistency
Checking…Versioned human review
Checking…Private JSON and TXT exports
Checking…ViciousByte evidence signature
Checking…Image OCR recognition
OCR not implementedGeneric table extraction
Table extraction not implementedAI-based extraction
AI extraction not implementedFiscal authenticity certification
Does not verify fiscal authenticity
What each result means
Bounded schema: classification, issuer RFC, receiver RFC, UUID, issue date, currency, subtotal, discount, transferred tax, withheld tax and total. This does not extract every template or every table in a document.
For PDF, candidates cite the page and the exact extracted text span, with Unicode code-point offsets; these are not image coordinates. XML preserves XPath, attribute and quoted value. The dashboard offers private plain text without executing or presenting active file content.
Confidence is ordinal: exact, candidate, ambiguous or unknown; it is not an accuracy percentage. exact for XML means the attribute was extracted, not that it is authentic. unknown means this method did not extract it, not that the document lacks it.
Value crosschecks report match, mismatch or unknown; arithmetic checks report consistent, inconsistent or unknown. An unknown result never becomes approval. Original checks do not certify a later manual correction.
Review covers all eleven fields, including classification. A correction with a note is recorded as a human assertion (human_asserted), separately from original candidates and without invented citations. Fields marked unknown remain unresolved. Exportable evidence preserves extraction, checks and review history.
Published formats, capacity and budget
Upload your own files only: no URLs or external providers are queried. XML must be UTF-8 CFDI 4 without DTDs or entities; encrypted PDF is not accepted. Image-only PDF or a document without extractable text retains that limitation and unknown fields; OCR is not used to fill gaps.
Document API uses its own document_attempt units, separately from ASR minutes, fiscal quotas and other services, through a pilot grant or paid project period. Each job includes its optional comparison XML in one attempt; a started local attempt may consume the unit even for invalid or unsupported content. Proposed plans and packs are shown subject to review; purchasing requires approval; there are no automatic charges.
Current limits could not be confirmed. Check the catalog or dashboard before uploading; capacity, retention and permission are not assumed.
Private data and bounded access
Original files, extracted text and evidence belong to the same account and project, with authenticated access even for administrators. Process only documents you are entitled to use. This phase does not send them to an AI service or an external fiscal verifier.
Deleting a job or reaching its retention expiry blocks reads and new publication, then queues originals and derived data for purge. Physical deletion is asynchronous; retained slots are released after actual cleanup. Requesting cancellation during processing is not confirmation that cancellation has completed.
Document API requires a pilot grant or a current paid period for your account and project, with its own budget. Plans for other services do not enable it. Requesting access records pending interest for review; it grants no permissions and starts no payment.
Plans and additional units
Prepaid 30-day base per project and service. Manual renewal after expiry; no automatic charges or automatic overage purchases.
Checking commercial terms…
Proposed pricing
A public view of a commercial proposal that has not been approved. These amounts are not available purchase offers and do not enable charges. Final terms may change after review.
Proposal dated October 25, 2026 · USD · base amounts exclude VAT (16%) · 30 days · manual renewal · no automatic chargesStarter
USD 9.00
10 document attempts
Project base for one 30-day periodPro
USD 25.00
30 document attempts
Project base for one 30-day periodBusiness
USD 59.00
90 document attempts
Project base for one 30-day period10 document attempts
USD 9.00 · Proposed · subject to review
30 document attempts
USD 24.00 · Proposed · subject to review
60 document attempts
USD 42.00 · Proposed · subject to review
Request pilot access
Register a Document API request for your account. This does not create a subscription, activate permissions or execute a payment.