Anvia

OCR ​

OCR recovers text and structure from scans, images, and PDFs. The v1 OCR adapter is provided by @anvia/mistral rather than a provider-neutral core contract.

For documents with reliable selectable text, prefer text extraction and avoid unnecessary OCR cost.

1. Create the model ​

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

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

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

Pass an explicit model ID to ocrModel(modelId) when the application must pin a version.

2. Process uploaded bytes ​

ts
const result = await ocr.ocr({
  source: {
    type: 'bytes',
    data: upload.bytes,
    filename: upload.filename,
  },
  tableFormat: 'markdown',
  includeImageBase64: false,
})

console.log(result.markdown)

Byte sources are uploaded through the Mistral files API before OCR. result.uploadedFile records normalized upload metadata when that path is used. The adapter does not delete the provider-side file; apply the retention policy your application requires.

3. Use URL or existing-file sources ​

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

Other source variants are image_url and file_id. A URL must remain reachable by the provider for the request duration; prefer a short-lived signed URL.

4. Preserve page provenance ​

result.text and result.markdown contain the combined page Markdown. Use result.pages for page-level review or knowledge ingestion:

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

Normalized pages may also include images, tables, hyperlinks, header, footer, dimensions, and confidence data when returned by the provider.

Keep includeImageBase64 disabled unless a later stage needs extracted images. Verify totals, dates, identifiers, and signatures before product writes. OCR neither proves authenticity nor grants access to the source.

Next, return media through tool results.

Built for Anvia.