Nanonets OCR

@Nanonets OCR list all my OCR models and show me which ones are ready to extract invoices or receipts

The account has 4 OCR models configured. Two are production-ready: 'Invoice Extractor v2' (model ID: inv_9x3k) extracts vendor name, invoice number, line items, and totals; 'Receipt Parser' (model ID: rcp_4m1a) extracts merchant, date, and amount. 'Expense Report Beta' is still training (78% accuracy). 'Purchase Order Draft' has no training data yet. Both ready models support table extraction for line-item details.

This example shows how the List OCR Models tool surfaces model readiness and extraction scope before you commit a document. Useful when you inherit a Nanonets account or need to route different document types to different models. Requires a valid API key with read access to all models in the workspace.

@Nanonets OCR extract all fields from this invoice PDF https://example.com/invoice-2024-Q4.pdf using my invoice model

Processed 3 pages from invoice-2024-Q4.pdf (file ID: file_7h2p). Page 1 extracted: Vendor='Acme Supplies Inc.', Invoice Number='INV-2024-1847', Date='2024-12-15', Total='$4,287.50', and a 12-row line-item table with descriptions, quantities, and unit prices. Pages 2–3 contained terms and conditions; no additional fields extracted. Confidence scores range from 0.91 to 0.98 for all extracted values.

The Extract From Public URL tool runs synchronously, so large multi-page documents may time out. Nanonets returns durable file and page IDs even when confidence is low or no fields match the model's schema. You can later call Get OCR Prediction Page to retrieve per-page details or re-check extraction after model retraining.

@Nanonets OCR pull the last 30 days of receipt predictions from my expense model and draft a spend summary by merchant category

Retrieved 47 receipt predictions from the past 30 days (model: rcp_4m1a). Top spending categories: Meals & Entertainment $1,823 (18 receipts), Office Supplies $1,104 (12 receipts), Travel & Lodging $987 (9 receipts), Software Subscriptions $640 (5 receipts), Miscellaneous $203 (3 receipts). Total extracted spend: $4,757. Note: 4 receipts had low-confidence merchant names and were grouped under Miscellaneous. Consider reviewing file IDs file_9k3x, file_2p7n for manual validation.

This synthesis example pairs List OCR Predictions (which fetches compact summaries over a date range) with the AI's categorization logic. Nanonets returns all matching predictions without pagination, so narrow your date range if the model processes hundreds of documents daily. The AI flags low-confidence extractions, prompting human review before financial reconciliation.

When Nanonets OCR beats manual invoice entry at small scale

A 3-person accounting team processing 40-80 vendor invoices a week hits the sweet spot for this MCP. One person trains a Nanonets model on the five invoice layouts they see most often—vendor name, date, line items, total—then the team uses Switchy to extract fields from new PDFs as they arrive in Slack or email. The Extract From Public URL tool pulls structured data in under 10 seconds, and the team reviews extracted fields in the same chat thread where they discuss payment approval. This works because Nanonets handles layout variance better than regex scripts, and 80 invoices a week justifies the API cost without needing a full AP automation platform. If your invoice volume drops below 20 a week or you need real-time processing inside an ERP, the setup overhead outweighs the time saved.

Using OCR predictions to route support docs faster

A 6-person support team receives 200 tickets a day, and 30 percent include scanned forms, screenshots, or PDFs that need manual review before routing. They configure a Nanonets model to extract claim numbers, account IDs, and issue categories from the five most common document types, then use Switchy to run Get OCR Prediction File on new attachments during ticket intake. The MCP returns structured fields within seconds, and the team uses those fields to tag tickets and assign them to the right specialist without opening every PDF. This cuts first-response time by 15 minutes per ticket when the document matches a trained layout. The trade-off: if more than half your attachments are one-off formats or handwritten notes, Nanonets accuracy drops below 70 percent and manual review becomes faster than fixing bad extractions.

When Nanonets OCR makes sense for recurring contract review

A 2-person legal ops team reviews 15-25 vendor contracts a month, each 8-12 pages, looking for liability caps, termination clauses, and renewal terms. They train a Nanonets model on their standard contract template and three common vendor formats, then use Switchy to extract those clauses from new PDFs before redline review. The Get OCR Model tool shows which fields the model reliably finds, and the team uses List OCR Predictions to audit extraction accuracy over the prior 30 days. This setup works because the contract formats repeat and the team needs structured data for a compliance dashboard, not just search. If your contracts are one-off MSAs with no layout consistency, or you process fewer than 10 a month, a paralegal with Ctrl+F is faster and cheaper than training and maintaining an OCR model.