Anvia

Appearance

Mistral OCR ​

Type: Recipe

Outcome ​

Extract page Markdown from a remote PDF with Mistral OCR. Use OCR for scans and images without reliable selectable text; use an ordinary loader for text-based documents when possible.

Prerequisites ​

  • pnpm add @anvia/mistral @anvia/core
  • A server-side MISTRAL_API_KEY
  • An authorized URL that Mistral can fetch

Implementation ​

ts
import { MistralClient } from '@anvia/mistral'

const apiKey = process.env.MISTRAL_API_KEY
if (!apiKey) throw new Error('Set MISTRAL_API_KEY.')

const ocr = new MistralClient({ apiKey }).ocrModel()
const result = await ocr.ocr({
  source: {
    type: 'document_url',
    url: 'https://example.com/authorized-invoice.pdf',
    documentName: 'invoice.pdf',
  },
  tableFormat: 'markdown',
  extractHeader: true,
  extractFooter: true,
})

console.log('pages:', result.pages.length)
console.log(result.markdown)

For local bytes, use { type: 'bytes', data, filename }; the adapter uploads the file for OCR and returns upload metadata as uploadedFile.

Run and expected behavior ​

Run pnpm tsx mistral-ocr.ts. The normalized response contains combined text and markdown, page entries, optional model and usage metadata, and the raw provider response. OCR quality varies with scan clarity and layout.

Boundaries ​

Validate authorization, URL scheme and host, file size, MIME type, page count, and malware status. Use short-lived signed URLs and avoid embedded image base64 unless required because it can greatly increase response and trace size. OCR is not proof of document authenticity; verify totals, dates, identifiers, and signatures before writes.

In production, use durable workers for large documents, track and clean up uploaded provider files under your retention policy, encrypt stored results, restrict trace payload capture, and keep a human review path for critical fields.

Source and extensions ​

The current source contract and tested sources are in MistralOcrModel and its OCR tests. Next, process selected pages, upload bytes, or feed reviewed Markdown into retrieval.

Built for Anvia.