OCR models

OCR extracts text and document structure from scanned files and images. In v1, OCR is a provider-specific capability exposed by @anvia/mistral rather than a generic @anvia/core model helper.

1. Install and create the OCR model

The Mistral adapter releases independently from Core. Check its declared Core dependency range when upgrading.

pnpm add @anvia/core @anvia/mistral

import { MistralClient } from '@anvia/mistral'

const apiKey = process.env.MISTRAL_API_KEY

if (!apiKey) {
  throw new Error('MISTRAL_API_KEY is required')
}

const client = new MistralClient({ apiKey })

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

ocrModel() requires a model ID; use an ID supported by the current Mistral OCR API.

2. Process a document URL

The source describes where Mistral should read the document.

import { ocrModel } from './models'

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

console.log(result.markdown)

Supported sources include document URLs, image URLs, existing Mistral file IDs, and byte uploads.

3. Upload bytes for OCR

Byte sources are uploaded through the Mistral file API before OCR runs.

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

const invoice = await readFile('invoice.pdf')

const result = await ocrModel.ocr({
  source: {
    type: 'bytes',
    data: invoice,
    filename: 'invoice.pdf',
    visibility: 'user',
  },
  extractHeader: true,
  extractFooter: true,
})

The response includes uploaded-file metadata when the adapter performs an upload.

4. Read normalized pages

The adapter combines page Markdown while preserving page-level structure.

console.log(result.text)
console.log(result.markdown)

for (const page of result.pages) {
  console.log(page.index, page.markdown)
}

Optional response fields include model, usage information, annotations, uploaded-file metadata, and the raw provider response.

5. Protect documents

Validate source access, file type, size, page selection, and retention before OCR. Keep raw files in application-owned storage, preserve stable source metadata for citations, and authorize extracted text before passing it to agents or retrieval indexes.

Continue to Completions for direct text generation workflows.