Extract structured data with source evidence

Extract Data applies a reusable or user-defined schema to PDF, JPG/JPEG, PNG, and DOCX documents. Every machine-extracted value stays connected to the source document, page, text, and region when geometry exists. Missing or conflicting facts remain missing or ambiguous instead of being guessed.

Before you start

A signed-in account is required. Use only documents you are authorized to process. Keep the authoritative originals.

Best for

• Invoice fields and line-item tables.
• Contract references, dates, parties, and obligations.
• Bank-statement balances and transaction tables.
• Identity-document text extraction without identity-verification claims.
• Mixed batches that need the same deterministic output schema.

How to use it

1. Add one or more supported documents within the displayed batch limits.
2. Choose the built-in invoice, contract, bank-statement, or identity-document template, or define your own fields.
3. Save a browser-local template version when the schema should be reused.
4. Start extraction and follow the protected queue status.
5. Review every missing, ambiguous, validation-failed, or OCR-derived result.
6. Select a field or table cell to inspect its page, raw source text, origin, and normalized region when available.
7. Correct values only after checking the source. Corrections remain marked as user edits and retain the original evidence.
8. Review conservative multi-page table continuation and reject a merge when the pages do not belong together.
9. Mark review complete, then export XLSX, CSV, or JSON.

Processing and queue

One extraction batch is one logical customer action using the existing OCR page tariff. Native parsing, local OCR, table analysis, review, correction, and export do not create separate extraction charges. Queue idempotency protects a replay of the same submission.

Confidence and evidence

High, review recommended, review required, missing, ambiguous, and validation failed are deterministic review states, not calibrated probability claims. A region is shown only when the parser or OCR engine supplied geometry. DOCX can preserve page and text evidence while explicitly reporting that region geometry is unavailable.

Tables

Table cells preserve row and column coordinates, header status, merged row/column spans, raw text, normalized values, and source anchors. Multi-page continuation requires adjacent pages, repeated headers, aligned columns, compatible value types, and compatible context. Matching column counts alone are never enough.

Exports

• XLSX: Fields, logical table sheets, and an Evidence sheet with safe sheet names and typed scalar cells where reliable.
• CSV: UTF-8, RFC-style escaping, explicit field/table rows, and source columns.
• JSON: Stable complete extraction, review, template, evidence, correction, and table structure.

Text beginning with spreadsheet formula characters is neutralized in CSV and XLSX. No internal temporary path or provider credential is included.

Privacy and limitations

PDF/image OCR and document parsing run on self-hosted ALTABRIG infrastructure with local executables. No external OCR or generative-AI provider receives the document. Temporary parser files are removed after the job. Saved custom templates remain in the current browser under an account-scoped key.

Deterministic extraction works best with explicit labels and structured tables. OCR can misread source text, implicit DOCX pagination can be unavailable, handwriting is not reliable, and a result is not legal, financial, KYC, or authenticity verification. Always review important facts against the original.

Troubleshooting

Malformed, encrypted, macro-bearing, mismatched, oversized, excessive-page, excessive-pixel, or unsafe DOCX packages fail closed. A failed document in a safe mixed batch is reported separately so successful documents remain reviewable. Retry only when the queue exposes an eligible retry.

Open Extract Data

Related

• /help/enhance-pdf
• /help/document-compare
• /help/privacy-security