OCR

Mistral OCR extracts Markdown and page structure from PDFs and images. It is a separate provider model, not an input mode on completionModel(...).

import { MistralClient } from '@anvia/mistral'

const mistral = new MistralClient({
  apiKey: process.env.MISTRAL_API_KEY!,
})

const ocr = mistral.ocrModel({
    modelId: 'mistral-ocr-latest'
})

ocrModel({ modelId }) requires an explicit OCR model ID.

Process a document URL

const result = await ocr.ocr({
  source: {
    type: 'document_url',
    url: signedDocumentUrl,
    documentName: 'invoice.pdf',
  },
  tableFormat: 'markdown',
  includeImageBase64: false,
})

console.log(result.markdown)

The provider must be able to reach the URL for the duration of the request. Prefer a short-lived signed URL over making a private document permanently public.

For an image URL, change the source shape:

const result = await ocr.ocr({
  source: {
    type: 'image_url',
    url: signedImageUrl,
  },
})

Process local bytes

import { readFile } from 'node:fs/promises'

const result = await ocr.ocr({
  source: {
    type: 'bytes',
    data: await readFile('./invoice.pdf'),
    filename: 'invoice.pdf',
    expiry: 3600,
    visibility: 'user',
  },
  pages: [0, 1],
  includeImageBase64: false,
})

For byte sources, the adapter first uploads the file with the ocr purpose, then processes the returned file ID. Empty bytes and an empty filename are rejected. result.uploadedFile contains the upload ID and available metadata.

The adapter does not delete the provider file. Choose expiry and visibility, and implement any required cleanup according to the document's sensitivity.

To reuse an existing Mistral file, pass its ID directly:

const result = await ocr.ocr({
  source: {
    type: 'file_id',
    fileId: 'file_123',
  },
})

Use the normalized result

The result provides:

• text and markdown, both containing the page Markdown joined with blank lines
• pages, with each page's index, Markdown, images, and available structural metadata
• optional model, usageInfo, and documentAnnotation
• optional uploadedFile for a byte upload
• rawResponse for narrow diagnostics

Preserve page boundaries when downstream citations or review need provenance:

const pages = result.pages.map((page) => ({
  id: `${documentId}#page=${page.index}`,
  text: page.markdown,
  pageIndex: page.index,
}))

Available request controls include page selection, extracted-image limits, minimum image size, bounding-box and document annotations, header and footer extraction, table format, and confidence-score granularity. Use providerOptions only for provider fields not represented by the normalized request. It cannot replace the adapter's model or document.

Leave includeImageBase64 disabled unless a later step needs extracted images; base64 content can significantly increase response, memory, and trace sizes.

OCR can contain recognition errors. Validate totals, dates, identifiers, and other critical values before product writes. Continue to the broader OCR guide for choosing between OCR, text extraction, and model-visible media.