[Enhancement]: Adding azure mistral ocr endpoint

[arthurolivierfortin]

Support for Azure Mistral OCR Integration

For the centralized workspace, I would like to add support for using the Azure Mistral OCR endpoint.

Since Azure's Mistral OCR endpoint does not accept file uploads, the file stream must be passed directly in the performOCR request. This means the logic must skip the uploadDocumentToMistral step entirely.

I'm not yet sure how custom OCR integrations will be handled in the long term, so this implementation may evolve. For now, here’s the temporary solution I added using a constant defined in librechat.yaml.

---

File: api/server/services/Files/MistralOCR/crud.js

Line 145 — Flag to detect Azure endpoint usage

const usingAzureEndpoint = ocrConfig.usingAzEndpoint || false;

---

Line 172 — Conditionally skip uploadDocumentToMistral

const mistralFile = usingAzureEndpoint
  ? null
  : await uploadDocumentToMistral({
      filePath: file.path,
      fileName: file.originalname,
      apiKey,
      baseURL,
    });

---

Skip getSignedUrl if using Azure

const signedUrlResponse = usingAzureEndpoint
  ? null
  : await getSignedUrl({
      apiKey,
      baseURL,
      fileId: mistralFile.id,
    });

---

Convert file to Base64 and use data URL for Azure

let ocrUrl;

if (usingAzureEndpoint) {
  const buffer = await fasync.readFile(file.path);
  const base64 = buffer.toString('base64');
  ocrUrl = `data:image/jpeg;base64,${base64}`;
} else {
  ocrUrl = signedUrlResponse.url;
}

---

Call performOCR with the appropriate parameters

const ocrResult = await performOCR({
  apiKey,
  baseURL,
  model,
  url: ocrUrl,
  documentType: usingAzureEndpoint ? 'document_url' : documentType,
});

---

I’d like to refactor this logic in the future to align with upcoming custom OCR support, including avoiding assignments of null to mistralFile and signedUrlResponse, and conditionally skipping those blocks entirely.

That said, before moving forward, I’d like to know whether this feature is something you’d be interested in. If so, I’d appreciate your input on how you’d prefer this integration to be structured in order to align with future plans for custom OCR implementations.

[andreoav]

Looking forward to this feature. We have a Mistral AI Deployment on our azure account and we're planning to integrate with out LibreChat instance.