import os
from openai import OpenAI
from dotenv import load_dotenv

load_dotenv()

SYSTEM_PROMPT = """
You are a senior banking AI advisor for a high-end lending firm. 
Your goal is to provide a very professional, encouraging, and data-driven explanation for a loan decision.

Rules:
1. Divide your explanation into exactly 3 or 4 clear, distinct points.
2. Prefix EVERY point with the delimiter: @@POINT@@
3. Keep each point concise (1-2 sentences).
4. Do NOT use markdown headers like ###. Use **Bold** for emphasis.
5. Do NOT invent new financial data. Only use what is provided in the packet.
6. If the loan is rejected, ensure the points explain the 'Why' and refer to the specific 'Actionable Step' provided in the deterministic packet.

Format:
@@POINT@@ [First insight]
@@POINT@@ [Second insight]
...etc
"""

# Prioritized models for fallback logic
LLM_MODELS = [
    "mistralai/mistral-large-3-675b-instruct-2512",
    "meta/llama-3.1-405b-instruct",
    "mistralai/mixtral-8x7b-instruct-v0.1"
]

def generate_explanation(packet: dict) -> str:
    api_key = os.getenv("NVIDIA_API_KEY")
    
    # Check if we should use Deterministic Mode
    result_label = "Approved" if packet.get('prediction') == 'Y' else "Rejected"
    confidence_val = packet.get('confidence', 0) * 100
    dti_val = packet.get('dti_ratio', 0)
    suggestion = packet.get('optimized_suggestion', '')

    default_narrative = f"""
### **AI Advisor Analysis (Deterministic Mode)**
*Note: This is an automated diagnostic analysis.*

Based on our core financial engine, your application has been **{result_label}** with a confidence score of **{confidence_val:.2f}%**. 

**Key Insights:**
- Your current **Debt-to-Income (DTI)** ratio is **{dti_val:.2f}%**. 
- **Actionable Step:** {suggestion}
"""

    if not api_key or "YOUR_KEY" in api_key:
        return default_narrative
        
    client = OpenAI(
        base_url="https://integrate.api.nvidia.com/v1",
        api_key=api_key
    )

    user_content = f"""
    Result: {packet.get('prediction', 'Unknown')}
    Confidence: {packet.get('confidence', 0):.2f}%
    DTI Ratio: {packet.get('dti_ratio', 0):.2f}%
    Primary Suggestion: {packet.get('optimized_suggestion', 'None')}
    """

    # Multi-model Fallback Loop
    for model_name in LLM_MODELS:
        try:
            print(f"Attempting inference with model: {model_name}...")
            completion = client.chat.completions.create(
                model=model_name,
                messages=[
                    {"role": "system", "content": SYSTEM_PROMPT},
                    {"role": "user", "content": user_content}
                ],
                temperature=0.15,
                max_tokens=1024
            )
            
            content = completion.choices[0].message.content
            # Basic validation that model followed instructions
            if "@@POINT@@" in content:
                print(f"Success with model: {model_name}")
                return content
            else:
                print(f"Model {model_name} failed to provide structured output. Trying next...")
                
        except Exception as e:
            print(f"Error calling {model_name}: {e}. Retrying with fallback...")

    # Final Fallback if all AI models fail
    print("CRITICAL: All AI models failed. Returning deterministic narrative.")
    return default_narrative