Invoice OCR API that reads every line, in any layout

Each value comes with a confidence score, so you can measure accuracy on your own invoices and decide what goes straight through and what goes to review.

Yes. Mindee returns the total net amount, the total tax and the total amount as separate fields, along with a tax breakdown by rate (rate, taxable base and tax value) when several VAT rates appear on the same invoice.

Each value comes with a confidence score. Your team can therefore automatically approve invoices where totals reconcile (net + tax = total) and send only inconsistent ones to manual review. This is especially useful for VAT recovery and multi-rate invoices common in the EU.

General-purpose LLMs can read an invoice, but they aren’t built for production document processing. Their output can change from one run to the next, they sometimes “hallucinate” plausible but wrong values (a total that doesn’t exist, a mixed-up date), and they don’t tell you how reliable each field is.

‍

Mindee is purpose-built for this task. It always returns the same structured JSON schema, it gives a confidence score for every field, and it provides the position of each value on the page so it can be verified. This makes extraction predictable, auditable and easy to automate. Mindee is also faster and more cost-effective at high volume, because you don’t need prompt engineering or output parsing. You can check our Mindee vs LLMs comparison.

Yes. Mindee can support invoices from suppliers worldwide, so it handles different layouts, languages, date formats and number formats (1,234.56 vs 1.234,56) without any templates or configuration.

‍

Amounts are returned as normalized numbers, together with the detected currency code, so you can feed them directly into your ERP or accounting system. Supplier details such as company registration and VAT numbers are also extracted, which makes it easier to match invoices to vendors across countries.

A typical workflow has four steps. Invoices arrive by email, upload or scan. They are then sent to the Mindee API, either through an SDK (Python, Node.js, Java, .NET…) or through a no-code tool like Zapier, Make or n8n. Mindee returns structured JSON in a few seconds. Fields above your confidence threshold are pushed straight to your ERP, and low-confidence fields are routed to a person for review. Most teams start with a no-code prototype in an afternoon, then move to the API for production volume.

Yes, your data is highly secured and protected during the entire conversion process. As an enterprise-grade OCR solution, Mindee takes data privacy very seriously. Here are the key security measures in place:

• Certified compliance: Mindee's infrastructure is SOC 2 Type II certified and fully GDPR compliant, ensuring strict industry standards for data protection.
• Encryption: All documents and extracted data are fully encrypted both in transit (during the API call) and at rest.
• Data localization: You have control over where your data is processed, with the option to host and process your data on either EU or US servers.
• Strict privacy: Mindee respects data confidentiality. Your processed documents are not shared with third parties, and your private data is not used to train global AI models.

In short, your sensitive documents are processed in a secure, isolated environment and remain completely under your control.

We provide a 14 days free trial so you can fully test any of our OCR models — no credit card required. After this, we offer different pricing tiers depending of the volume of pages processed and the features you might need. See the pricing page for more information.

The supported document formats are:

• PDF (application/pdf) – While technically a document format rather than an image, it is natively supported for all extraction models.
• JPEG / JPG (image/jpeg)
• PNG (image/png) – Must be non-animated.
• WebP (image/webp)
• TIFF / TIF (image/tiff) – Both single-page and multi-page TIFFs are supported and processed similarly to PDFs.
• HEIC (image/heic) – Apple's High-Efficiency Image Container format.

‍

📌 Important technical limitations to keep in mind:

To ensure the OCR pipeline runs smoothly and successfully returns your JSON, make sure your files adhere to the following constraints:

• Maximum file size: 100 MB per file.
• Maximum page count: Up to 200 pages per document.
• File state: The files cannot be encrypted, and PDFs must not be password-protected.