Structured Outputs with Pydantic

Raw LLM text is unpredictable. Structured outputs give you typed, validated Python objects.

The pattern:

1. Define a Pydantic model for the expected shape
2. Instruct the LLM to respond with JSON matching that shape
3. Parse with Model.model_validate_json(response)

Why care? This is how production AI systems stay reliable. Instead of parsing strings, you get objects with guaranteed types and validation errors when the model goes off-script.

Task

Complete classify_expense(description, client) to return an ExpenseResult.

The ExpenseResult model and CATEGORIES list are already defined. Your system prompt must:

• Tell the model the allowed categories
• Request JSON with the shape {"category": "...", "confidence": 0.0, "reason": "..."}