mistral-ocr（米斯特拉尔 OCR）

Mistral OCR

Extract text from images and PDFs using Mistral's dedicated OCR API. No external dependencies required.

Requirements

This skill requires a Mistral API key. If you don't have one, follow the guide in reference/getting-started.md.

API Key

The user must provide their Mistral API key. Ask for it if not available.

Option 1 (Recommended for AI agents): User provides key directly in message:

"Use this Mistral key: aBc123XyZ..."
"Convert this PDF to markdown, my API key is aBc123XyZ..."

Option 2: Environment variable $MISTRAL_API_KEY

Option 3: Claude Code settings (~/.claude/settings.json)

If no key is available, guide the user to get one at console.mistral.ai.

API Endpoint

Use the dedicated OCR endpoint for all document processing:

POST https://api.mistral.ai/v1/ocr

Model: mistral-ocr-latest

Features

1. PDF → Markdown (Direct, no conversion needed!)

curl -s "https://api.mistral.ai/v1/ocr" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-ocr-latest",
    "document": {
      "type": "document_url",
      "document_url": "https://example.com/document.pdf"
    }
  }'

2. Image → Text

Works with JPG, PNG, WEBP, GIF:

curl -s "https://api.mistral.ai/v1/ocr" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-ocr-latest",
    "document": {
      "type": "image_url",
      "image_url": "https://example.com/image.jpg"
    }
  }'

3. Local Files (Base64 Data URL)

For local PDFs or images, encode as base64 and use a data URL.

ALWAYS use curl (works on all platforms including Windows via Git Bash):

# For local PDF
BASE64=$(base64 -w0 document.pdf)
curl -s "https://api.mistral.ai/v1/ocr" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-ocr-latest",
    "document": {
      "type": "document_url",
      "document_url": "data:application/pdf;base64,'"$BASE64"'"
    }
  }'

# For local images (PNG, JPG, etc.)
BASE64=$(base64 -w0 image.png)
curl -s "https://api.mistral.ai/v1/ocr" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-ocr-latest",
    "document": {
      "type": "image_url",
      "image_url": "data:image/png;base64,'"$BASE64"'"
    }
  }'

MIME types:

• PDF: data:application/pdf;base64,...
• PNG: data:image/png;base64,...
• JPG: data:image/jpeg;base64,...
• WEBP: data:image/webp;base64,...

4. Structured JSON Output

For invoices, forms, tables - ask for JSON in a follow-up or use Document AI annotations.

Response Format

The API returns markdown directly:

{
  "pages": [
    {
      "index": 0,
      "markdown": "# Document Title\n\nExtracted content here...",
      "images": [],
      "tables": [],
      "dimensions": {"dpi": 200, "height": 842, "width": 595}
    }
  ],
  "model": "mistral-ocr-latest",
  "usage_info": {"pages_processed": 1, "doc_size_bytes": 12345}
}

Workflow

User requests OCR from image or PDF

1. Get API key - Ask user if not in environment
2. Determine input type (URL or local file)
3. For local files, ALWAYS use temp file approach (avoids "Argument list too long" error):

# Cross-platform temp directory
TMPDIR="${TMPDIR:-${TEMP:-/tmp}}"

# Step 1: Encode file to base64
base64 -w0 "document.pdf" > "$TMPDIR/b64.txt"

# Step 2: Create JSON request file
echo '{"model":"mistral-ocr-latest","document":{"type":"document_url","document_url":"data:application/pdf;base64,'$(cat "$TMPDIR/b64.txt")'"}}' > "$TMPDIR/request.json"

# Step 3: Call API with -d @file (use actual key, not variable)
curl -s "https://api.mistral.ai/v1/ocr" \
  -H "Authorization: Bearer YOUR_API_KEY_HERE" \
  -H "Content-Type: application/json" \
  -d @"$TMPDIR/request.json" > "$TMPDIR/response.json"

# Step 4: Extract markdown with node (NOT jq - not available on all systems)
node -e "const fs=require('fs'); const r=JSON.parse(fs.readFileSync('$TMPDIR/response.json')); console.log(r.pages.map(p=>p.markdown).join('\n\n---\n\n'))"

1. Save to.md file using Write tool
2. Confirm file location to user

IMPORTANT: Cross-Platform Compatibility

• ALWAYS use curl (works on Windows via Git Bash)
• ALWAYS use -d @file for request body (handles large files)
• NEVER use jq - use node instead to parse JSON
• Use ${TMPDIR:-${TEMP:-/tmp}} for temp files (works on all systems)
• Copy response.json to user directory before parsing with node on Windows

Usage Examples

When the user says:

Error Handling

Supported Formats

No external dependencies required! Unlike other OCR solutions, Mistral OCR handles PDFs directly without needing pdftoppm, ImageMagick, or any other tools.

Pricing

As of 2025, Mistral OCR pricing:

• $2 per 1,000 pages
• 50% discount with Batch API

Check current rates at mistral.ai/pricing

References

• Getting Started - How to get your API key
• PDF to Markdown - PDF conversion examples
• Output Formats - JSON, Markdown, plain text
• Step-by-Step Guide - Complete tutorial with examples

*Skill by Parlamento AI*