app/llm_service.py

import logging
from typing import Optional
from llama_cpp import Llama
from app.config import get_settings

logger = logging.getLogger(__name__)
_llm_model: Optional[Llama] = None

def get_llm() -> Llama:
    global _llm_model
    if _llm_model is None:
        settings = get_settings()
        
        if not settings.model_path:
            raise ValueError("Model path is empty. Download failed?")

        logger.info(f"⚡ Loading Llama-3 from: {settings.model_path}")
        
        # High-Performance Configuration
        _llm_model = Llama(
            model_path=settings.model_path,
            n_ctx=settings.model_n_ctx,
            n_threads=settings.model_n_threads,
            n_parallel=settings.model_n_parallel, # CRITICAL for 1000 users
            n_batch=512, # Processes tokens in chunks (Faster)
            verbose=False
        )
        logger.info("✅ Brain Loaded with Parallel Powers")
    
    return _llm_model

ISLAMIC_SYSTEM_PROMPT = """You are Nur, a knowledgeable and compassionate Islamic assistant. 
Guidelines:
1. Base responses on Quran and Sahih Hadith.
2. Be respectful and gentle.
3. Cite sources clearly.
4. Begin with "Bismillah"."""

def generate_islamic_reflection(prompt: str, max_tokens: int = 512, temperature: float = 0.7):
    llm = get_llm()
    
    full_prompt = f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{ISLAMIC_SYSTEM_PROMPT}<|eot_id|><|start_header_id|>user<|end_header_id|>

{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
"""
    
    response = llm(
        full_prompt,
        max_tokens=max_tokens,
        temperature=temperature,
        stop=["<|eot_id|>", "<|end_of_text|>"],
        echo=False
    )
    
    return {
        "reflection": response["choices"][0]["text"].strip(),
        "usage": response["usage"]
    }