Contract analysis

Use process_document in core, metier-docs, or admin. Check mistral://capabilities for the current endpoint and tool availability.

Process the source

Prefer supplied contract text or Markdown:

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

For a PDF or image requiring OCR, replace source with { "type": "url", "url": "https://example.com/contract.pdf" } or { "type": "file_id", "fileId": "<existing Mistral file ID>" }.

For local files, use an available converter to obtain text, or upload document bytes with files_upload in admin: filename, content_base64, purpose: "ocr". Its structuredContent.id becomes source.fileId. Do not pass a filesystem path as a URL. If these routes are unavailable, request text or a provider-accessible source.

Text skips OCR and uploads, but contract extraction still sends text to the configured chat endpoint. OCR is optional and depends on account access and quota; no free-processing guarantee applies. The tool uses the configured chat default and has no model argument. Cache bypass affects the local extraction cache only; it does not establish provider retention or data residency.

Review the extraction

Check isError before consuming structuredContent. The contract payload contains:

• parties[]: name and optional nullable role.
• clauses[]: heading, text, and optional nullable risk (low, medium, or high).
• risk_score: a nullable number from 0 to 1.
• key_dates[]: label and iso.
• summary: nullable text.
• ocr_text, extraction_source, ocr_confidence, and page_count for provenance. The last two are null for supplied text.

Present the summary, parties, dates, and clauses with their returned risk labels. Put high-risk clauses first, then medium, low, and unassessed clauses. Ground each concern in the clause text and the user's context. Treat risk_score as a model assessment, not a calibrated legal probability. Do not add a critical level, invented missing-protection fields, or categorical rules about enforceability.

Typed extraction accepts up to 60,000 UTF-16 code units. OCR selects the first 50 pages by default, adjustable with options.maxPages up to 200. State coverage limits, preserve clause context when splitting long documents, and do not present partial analysis as a complete review. An OCR confidence error requires a clearer source or reviewed text, not a silently weakened threshold.

For a requested French summary, retrieve the MCP prompt french_legal_summary with legal_text and audience (juriste, dirigeant, or grand_public). Convert each returned text message to {role, content: message.content.text} before calling mistral_chat; omit model to honor the server default. Keep extracted facts distinct from interpretation and carry through any missing information.