deepread-ocr_skill

Overview

This skill turns scanned PDFs and images into high-accuracy structured data with minimal manual review. It uses multi-model consensus and multi-pass validation to deliver 97%+ accuracy, flagging only uncertain fields for Human-in-the-Loop (HIL) review. Zero prompt engineering is required and a free tier covers 2,000 pages/month.

How this skill works

Upload documents (PDF or image) to the API; the pipeline performs rotation correction, OCR, multi-model validation, and structured extraction. Results include cleaned text, JSON fields with confidence scoring, per-page quality flags, and preview URLs for HIL review. Uncertain fields carry hil_flag=true and an explanatory reason so automated workflows can auto-process confident data and route exceptions to human reviewers.

When to use it

• Automating invoice, receipt, or form data extraction
• Digitizing contracts and extracting dates, parties, and clauses
• Building document workflows where you must know which extractions need review
• Any application requiring high-accuracy OCR with minimal manual effort
• Projects where asynchronous processing (2–5 minutes) and webhooks are acceptable

Best practices

• Use webhooks in production to avoid polling and lower latency
• Design descriptive JSON schemas or reusable blueprints for consistent extraction
• Auto-process fields with hil_flag:false and route hil_flag:true items to a review queue
• Poll only for development or when you can’t expose a webhook; check every 5–10 seconds
• Create and reuse blueprints to improve accuracy for recurring document types

Example use cases

• Invoice processing: extract vendor, totals, and line items with confidence scores
• Receipt parsing for expense automation and bookkeeping
• Form digitization: turn paper intake forms into structured records
• Contract intake: capture parties, dates, and key clauses and flag uncertain fields
• Quality-critical apps where knowing which fields need human review reduces manual work to ~5–10%