OCR Disclosure

Why this matters

Many uploaded PDFs are scanned images without a native text layer. To turn those into searchable text we use an external OCR provider. This page documents what gets sent and when.

Routing

1. When you upload a PDF, we first try native text extraction without an external call.
2. If the native layer is missing or low-confidence, the file is sent to Mistral OCR for extraction.
3. We store provider name, model version, and per-page confidence on extraction records for auditability.

Both scopes

Public legal corpus uploads and user-private uploads route to Mistral OCR when native extraction is insufficient. A single provider keeps Arabic and English layout handling consistent.

What is sent

• The raw PDF bytes for pages requiring OCR.
• No account email, IP address, or user identifier.

What is NOT sent

• Image files skip OCR entirely.
• DOCX uploads are converted locally.
• Plain-text uploads pass through unchanged.
• PDFs with a usable native text layer skip OCR.

Provider policy

Mistral AI data-handling terms apply after processing. See https://mistral.ai/terms/ for provider terms.

On-prem option

A self-hosted OCR provider is on the roadmap for users who cannot send documents to third parties. Until then, upload PDFs that already have a text layer to avoid OCR.