Data Extraction

Data extraction is the platform's pipeline for pulling structured information out of unstructured or semi-structured documents — invoices, receipts, IDs, statements, forms. Instead of humans typing the data, the platform reads the document and populates fields automatically.

What Gets Extracted

• Form fields: Named fields on standard documents (invoice number, total, due date, supplier name).
• Tables: Line items with quantity, description, unit price, total.
• Identity data: Names, dates of birth, document numbers from IDs and passports.
• Signatures and stamps: Detected and cropped for verification workflows.
• Free-form fields: Custom “find this value near this label” rules.

How It Works

1. Document uploaded.
2. OCR runs if the document is image-only. Cloud OCR (AWS Textract, Google Document AI, Azure Form Recognizer) optional for higher accuracy.
3. Field-detection rules look for known labels and patterns.
4. Extracted values populated as a draft record with confidence scores per field.
5. Anything below the confidence threshold is queued for human review.
6. Final record saved and the source document attached.

Worked Examples

• Accounts payable: Supplier emails an invoice PDF. Extraction populates supplier, date, line items, totals. Anything below 90% confidence is reviewed by a clerk before posting.
• HR onboarding: New hire uploads scanned passport and ID. Extraction populates DOB, passport number, expiry date, name. Compliance check verifies expiry > 6 months ahead.
• Expense receipts: Employee photographs a restaurant receipt. Extraction captures merchant, date, amount, VAT, currency.
• Application forms: Scanned paper applications digitised — extraction populates structured fields ready for the existing workflow.

An extraction rule is a configuration that tells the platform what to extract from a specific kind of document. You create one per document type — supplier invoices, receipts, application forms, ID documents — each with its own field map.

Step-by-Step

1. Open Architect Panel → Data Extraction → Rules and click New Rule.
2. Upload a sample document to anchor on.
3. Name the rule (e.g. “Standard supplier invoice”) and pick the target datastore.
4. For each target field, define how to find it:
   • Label nearby: Look for a label (“Invoice #”) and read the value to its right or below.
   • Regex: Match a regex against the document text.
   • Bounding box: Fixed coordinates (great for highly-standardised forms).
   • ML model: Use a pre-trained model for the document type.
5. Set the confidence threshold (e.g. 85%).
6. Test against more sample documents. Tune rules until extraction is reliable.
7. Publish.

Auto-Selecting a Rule

When a new document arrives, the platform identifies which rule to use:

• By upload context (e.g. uploads through the “supplier invoice” form always use the invoice rule).
• By sender (emails from a specific supplier use a supplier-specific rule).
• By document classification (ML classifier picks among defined rules).

Worked Examples

• Supplier-specific invoice rule: One per major supplier (each has a different layout). Falls back to a generic rule for unknown senders.
• Receipt rule: ML model trained on common receipt layouts; extracts merchant, date, total, line items.
• ID document rule: Bounding-box rule for standard passports — MRZ zone parsed, photo cropped, expiry validated.
• Form rule: Fixed bounding-box rule for an internal application form printed from your own template.

The extraction pipeline handles a wide range of document types. This article lists what's supported, with notes on accuracy and tuning per type.

High-Accuracy Document Types

• Modern PDF invoices: Text-based PDFs from common accounting systems extract reliably. 95%+ accuracy on totals, dates, supplier names.
• Passport MRZ zone: Machine-readable zone yields near-perfect extraction.
• Pre-printed forms: Fixed-layout forms with checkbox marks and known field positions.
• Modern receipts: Thermal receipts photographed in good light read well.

Variable-Accuracy Types

• Photographed receipts in low light: Wrinkled, faded or dimly lit images degrade OCR.
• Handwritten forms: Tesseract struggles; cloud OCR is significantly better but still imperfect.
• Scanned old documents: Faded ink, marginalia, stamps confuse layout analysis.
• Multi-page invoices with continuation tables: Detecting where the table continues across pages requires careful rule design.

Limited Support

• Drawings and diagrams with embedded labels.
• Heavily formatted marketing PDFs where data is artistic rather than structured.
• Languages without Latin script in default Tesseract; cloud OCR handles most.
• Encrypted documents can't be read without decryption.

Worked Examples

• Accounts payable team: Receives invoices in mixed formats. PDF-from-Word ones extract perfectly; scanned ones go to review queue.
• Employee onboarding: Passport MRZ extraction works; photo extraction works; signature page requires manual confirmation.
• Multi-language receipts: Cloud OCR with language hints for tenants in Germany, Spain, Brazil.
• Multi-page contracts: Extract counterparties from the cover page; extract signature page details from the back; review middle pages manually.