app.py

import os, json, asyncio
from fastapi import FastAPI, Request, HTTPException
from fastapi.responses import StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
from prompts import build_system_prompt
from search_engine import search_web

app = FastAPI()
app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])

MODEL_REPO = "bartowski/Qwen_Qwen3.6-35B-A3B-GGUF"
MODEL_FILE = "Qwen_Qwen3.6-35B-A3B-IQ3_M.gguf"
llm = None

def load_model():
    global llm
    if llm is None:
        print("⬇️ جاري تحميل النموذج...")
        model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
        llm = Llama(
            model_path=model_path,
            n_ctx=2048,          # متوازن مع 18GB RAM (يمكن رفعه لـ 3076 إذا توفرت رامات إضافية)
            n_threads=4,
            n_gpu_layers=0,
            use_mmap=True,
            verbose=False
        )
        print("✅ تم تحميل النموذج بنجاح.")

@app.on_event("startup")
def startup():
    load_model()

def format_qwen_chat(messages: list, system_prompt: str) -> str:
    """بناء قالب محادثة Qwen3 الصحيح مع حفظ السياق"""
    prompt = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
    # نحتفظ بآخر 5 رسائل فقط لتوفير سياق الذاكرة على السيرفر المجاني
    history = messages[-5:] if len(messages) > 5 else messages
    for msg in history:
        role = "user" if msg["role"] == "user" else "assistant"
        prompt += f"<|im_start|>{role}\n{msg['content']}<|im_end|>\n"
    prompt += "<|im_start|>assistant\n"
    return prompt

async def generate_stream(messages: list, mode: str):
    system_prompt = build_system_prompt(mode)
    
    # وضع البحث: حقن النتائج بتعليمات واضحة
    if mode == "search":
        query = messages[-1]['content']
        search_res = search_web(query)
        # نضيف النتائج كرسالة نظام قبل آخر رسالة مستخدم
        messages = messages.copy()
        messages.insert(-1, {"role": "system", "content": f"[SEARCH RESULTS]\n{search_res}\n\nINSTRUCTION: Use the above results to answer accurately. If irrelevant, rely on your knowledge."})

    prompt = format_qwen_chat(messages, system_prompt)
    
    # إعدادات توليد محسنة لنماذج MoE الكبيرة
    for token in llm(
        prompt, 
        max_tokens=2048, 
        stop=["<|im_end|>", "<|user|>"], 
        stream=True, 
        temperature=0.7,
        repeat_penalty=1.1,  # منع التكرار
        top_p=0.9
    ):
        yield json.dumps({"token": token["choices"][0]["text"]}) + "\n"
        await asyncio.sleep(0.01)

@app.post("/v1/chat/completions")
async def chat_completions(request: Request):
    data = await request.json()
    messages = data.get("messages", [])
    mode = data.get("mode", "chat")
    if not messages:
        raise HTTPException(400, "No messages provided")
    return StreamingResponse(generate_stream(messages, mode), media_type="application/json")

@app.get("/health")
def health():
    return {"status": "ok", "model": MODEL_FILE}