Document & Invoice Extraction API

Extract Data From Documents & Invoices

Turn PDFs, scans, and photos into structured JSON—invoice numbers, amounts, dates, signatures—without custom OCR pipelines. Use unoblox's vision-capable models, billed in rupees.

Why vision over traditional OCR

Traditional OCR struggles with layouts, handwriting, and context. Vision models (Claude, Qwen) understand structure, read tables, and extract meaning. Faster iteration, fewer exceptions.

Models for document extraction

Model	Cost (input/output)	Best for	Context
Claude Opus	₹504 / ₹2520	High-accuracy legal, complex layouts	200K+ tokens
Claude Sonnet	₹201.6 / ₹1008	Fast extraction, most invoices	200K+ tokens
Qwen3.8-27B	₹16.32 / ₹48.96	Budget, open-weight	262K tokens
GPT-5	₹126 / ₹1008	Reasoning over ambiguous docs	400K tokens

3-step extraction workflow

Step 1: Upload document as base64

POST https://api.unoblox.ai/v1/chat/completions
Authorization: Bearer ub-gw-...

{
  "model": "anthropic/claude-opus-4-5",
  "messages": [
    {"role": "user", "content": [
      {"type": "text", "text": "Extract invoice number, date, amount, vendor name as JSON."},
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,...base64..."}}
    ]}
  ]
}

Step 2: Parse JSON response

{"invoice_number": "INV-2026-001", "date": "2026-09-19", "amount": "₹50,000", "vendor": "Acme Inc."}

Step 3: Validate and store — Schema-validate the JSON, retry on parse errors, store in your database.

Real-world pricing for 1,000 invoices

• Claude Sonnet: ~₹200–300 (depending on document length)
• Qwen3.8-27B: ~₹50–80
• No SaaS fees, no per-document charges, just token usage

Frequently asked questions

Q: How do I handle handwritten or poor-quality scans? A: Claude Opus excels at degraded images; Qwen3 is faster but less tolerant. Test both on your worst-case docs; fallback to Opus for edge cases.

Q: Can I extract tables from multi-page PDFs? A: Yes, if the PDF fits into token limits (~200k for Claude models). For very long documents, split into pages and batch-process.

Q: What about GSTIN, PAN, or bank account validation? A: The API extracts raw text; add post-processing to validate checksums (GSTIN, IFSC) or regex patterns.

Q: Do you store the documents I upload? A: No. unoblox and upstream providers process images stateless; no retention. See compliance docs for DPDP/GDPR details.

Q: Which format should I send—PDF or image? A: Turn PDFs into PNG or JPG first (using PyPDF2 or ImageMagick), then send as a base64 image. Simpler and more reliable than raw PDF parsing.

Q: Can I extract from handwritten forms? A: Yes, Claude Opus reads cursive well. Qwen3 is faster but less reliable on cursive. Test on a sample.

Get started in rupees → https://unoblox.ai/sign-in