AI Model Wiki

    Skip to main content
    20 min
    Lesson 5 of 5

    Hands-On Practice

    From the course Mistral AI: European Open Models

    Put what you have learned in this course into practice with the structured exercise below. Working through it end-to-end is the single best way to consolidate the concepts.

    Concrete Starter Project: Build a Multilingual Customer Support Classifier

    This project showcases what Mistral does best: European multilingual tasks, function calling, and structured output — all at low cost using Mistral Small.

    What You'll Build

    A support ticket classifier that reads incoming emails in any European language, detects the language, classifies the intent (billing, technical, general), assigns a priority, and routes to the correct team — all with a single Mistral Small call.

    Step 1 — Define the Classification Function

    from mistralai import Mistral
    import json
    
    client = Mistral(api_key="your-api-key")
    
    tools = [{
        "type": "function",
        "function": {
            "name": "classify_ticket",
            "description": "Classify a customer support ticket",
            "parameters": {
                "type": "object",
                "properties": {
                    "detected_language": {"type": "string", "description": "ISO 639-1 language code, e.g. 'fr', 'de', 'es'"},
                    "intent": {"type": "string", "enum": ["billing", "technical", "account", "general", "urgent"]},
                    "priority": {"type": "string", "enum": ["low", "medium", "high", "critical"]},
                    "summary_in_english": {"type": "string", "description": "Brief English summary of the issue"},
                    "suggested_team": {"type": "string", "enum": ["billing-team", "tech-support", "account-management", "general-support"]}
                },
                "required": ["detected_language", "intent", "priority", "summary_in_english", "suggested_team"]
            }
        }
    }]
    

    Step 2 — Classify Incoming Tickets

    def classify_support_ticket(ticket_text: str) -> dict:
        response = client.chat.complete(
            model="mistral-small-latest",  # Cost-effective for classification
            messages=[
                {
                    "role": "system",
                    "content": "You are a multilingual customer support classifier. Analyze tickets in any language and classify them accurately."
                },
                {
                    "role": "user",
                    "content": f"Classify this support ticket:\n\n{ticket_text}"
                }
            ],
            tools=tools,
            tool_choice="any"  # Force tool use for structured output
        )
    
        tool_call = response.choices[0].message.tool_calls[0]
        return json.loads(tool_call.function.arguments)
    
    # Test with multilingual tickets
    tickets = [
        "Bonjour, je ne peux pas accéder à mon compte depuis hier soir.",  # French
        "Hola, me han cobrado dos veces este mes en mi tarjeta.",           # Spanish
        "Mein Abonnement wurde ohne Vorwarnung gekündigt!",                 # German
        "I cannot connect to the API — getting 503 errors.",               # English
    ]
    
    for ticket in tickets:
        result = classify_support_ticket(ticket)
        print(f"Language: {result['detected_language']}, Intent: {result['intent']}, Priority: {result['priority']}")
        print(f"  Summary: {result['summary_in_english']}")
        print(f"  Route to: {result['suggested_team']}\n")
    

    Step 3 — Benchmark Against GPT-4o-mini

    Run the same 20 test tickets through both mistral-small-latest and gpt-4o-mini. Compare:

    • Classification accuracy (manually label ground truth first)
    • Average latency per request
    • Cost per 1000 tickets
    • Language detection accuracy across all five supported languages

    Mistral Small typically matches GPT-4o-mini on European language tasks and may outperform it on French and German specifically.

    Self-Check Questions

    • What happens when a ticket mixes two languages (common in Swiss/Belgian markets)?
    • How would you handle tickets that are ambiguous between technical and billing (e.g., "my API key was revoked after my payment failed")?
    • At what volume (tickets per month) does fine-tuning Mistral Small become cheaper than prompting?

    Stretch Goal

    Add a second Mistral call to draft a reply in the customer's detected language, using the English summary as input. Compare draft quality between Mistral Large and Mistral Small for the reply generation step.

    Key Takeaways

    • •Hands-on application converts course concepts into real skill
    • •Start small with an MVP before optimising
    • •Self-check questions catch gaps in understanding before they ship to production
    0 of 5 lessons marked complete

    Progress is kept in this browser only.