extract.do
Turn Unstructured Content into Structured Data
Extract tables, forms, invoices, and custom data from PDFs, images, and documents using AI. Convert to JSON, CSV, or custom schemas with automatic validation.
import { Extract } from 'extract.do' // Extract invoices from email attachmentson.Email.with.Attachment(async $ => { const invoice = await Extract.fromPDF({ file: $.Attachment.buffer, schema: $.InvoiceSchema }) await $.Database.insert.Invoice(invoice) if (invoice.confidence < 0.9) $.HumanReview.queue({ invoice })}) // Scrape competitor pricing hourlyExtract.every.hour(async $ => { const products = await Extract.fromHTML({ urls: $.Competitors.urls, parallel: 10 }) $.Analytics.track.CompetitorPricing(products)})
One of 48 primitives
extract.do sits in Agents: reason, research and act. Every .do shares one grammar, Business-as-Code, and one SDK.
Data
Records, vocabulary and search
Insight
Measure, test and trace
Frequently asked questions
Extract.do processes PDFs (including scanned documents via OCR), images (PNG, JPG, TIFF), HTML web pages, emails (with attachments), Word documents, spreadsheets, JSON/XML, and plain text. AI vision models handle tables, forms, handwriting, signatures, stamps, and complex layouts. Support for 50+ languages with automatic language detection. Process documents up to 100 pages or 50MB each.
95%+ accuracy on standard document types (invoices, receipts, forms) with confidence scores for every extracted field. Confidence combines OCR certainty, schema validation, and cross-field consistency checks. Use multi-model validation (GPT + Claude consensus) for critical extractions to reach 99%+ accuracy. Built-in detection of ambiguous, missing, or conflicting data with automatic flagging for human review.
Yes. Define schemas using TypeScript types, JSON Schema, or natural language descriptions. Support for nested objects, arrays, enums, date/time formats, regex patterns, and custom validation functions. Extract.do automatically validates extracted data against schemas, applies business rules (e.g., invoice total must match line items sum), and detects anomalies. Version control for schemas with backward compatibility.
Automatic image enhancement (deskew, denoise, contrast adjustment) before OCR. Multi-pass extraction with different AI models for difficult documents. Confidence thresholds trigger human review workflows—integrate with your existing review tools or use built-in review UI. Continuous learning: human corrections improve model accuracy over time. Fallback to manual processing for extremely low-quality documents.
Process 1000s of documents in parallel across 300+ edge locations. Typical extraction: <2 seconds per page for PDFs, <500ms for web pages, <1 second for images. Cost: $0.01 per document page, $0.001 per web page extraction. First 1000 pages/month free. Bulk discounts for enterprise volumes. Automatic rate limiting and queue management for large batches. Streaming results for real-time processing.