pdf-invoice-extractor

Invoice extraction

Use process_document, available in core, metier-docs, and admin. Check mistral://capabilities if it is unavailable. Prefer existing text or Markdown; request OCR only when the document still needs it.

Process the source

Call process_document with an explicit invoice kind and local cache bypass:

{
  "source": { "type": "text", "text": "<invoice text or Markdown>" },
  "kind": "invoice",
  "options": { "cache": "bypass" }
}

For OCR, replace source with one of:

• { "type": "url", "url": "https://example.com/invoice.pdf" } for a provider-accessible PDF or image.
• { "type": "file_id", "fileId": "<existing Mistral file ID>" } for an uploaded document.

A local path is not a source URL. If a local converter is available, use its extracted text. Otherwise files_upload requires admin and accepts filename, content_base64 containing the file bytes, and purpose: "ocr". Use its structuredContent.id as source.fileId. If neither route is available, request text, an accessible URL, or an existing upload ID.

Text input skips OCR and the Files API; invoice extraction still sends text to the configured chat endpoint. OCR requires account access and quota, and processing is not guaranteed to be free. Omit model selection: this tool uses the configured chat default internally. options.cache: "bypass" prevents this call from reading or writing the local extraction cache; it is not a provider retention setting.

Interpret the result

Check isError first. On success, use structuredContent with kind: "invoice":

Field	Meaning
vendor.name, optional nullable vendor.tax_id	Extracted vendor identity
total, currency	Nullable total and three-character currency code
line_items[]	desc, qty, unit_price, amount
due_date	Nullable due date
anomalies[]	Reported extraction anomalies
ocr_text, extraction_source	Text used and whether it came from provided_text or mistral_ocr
ocr_confidence, page_count	Both null for supplied text; OCR metadata otherwise