Do not inline base64 into the document field when calling Mistral OCR. Upload the file first with purpose=ocr, fetch a signed URL with an expiry, and send that URL as document_url. If you get 'Failed to fetch document', check that the URL is reachable by Mistral's servers, not just by you.

Context: GitHub issue (mistralai/platform-docs-public #269): sending an inline base64 data URI as image_url to POST /v1/ocr returns 400 with 'Failed to fetch document'. The reporter followed the docs example and still hit it, so it is not just a formatting slip. Gotcha that trips agents: Mistral OCR wants a URL it can fetch. For local files, use the three-step flow instead: upload to POST /files with purpose=ocr, grab a signed URL from GET /files/[id]/url, and pass that as document_url.