app.py

import os
import logging
from fastapi import FastAPI, Request
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse
from llama_cpp import Llama

# 1. Setup Logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# 2. Model Configuration
MODEL_PATH = "./models/gpt-oss-20b-Q3_K_M.gguf"
llm = None

# 3. Memory Storage
sessions = {}
MAX_HISTORY = 6  # Keep last 3 full turns

def load_model():
    global llm
    if llm is None:
        logger.info("🔥 Initializing 20B Harmony Engine...")
        try:
            llm = Llama(
                model_path=MODEL_PATH,
                n_ctx=2048,        
                n_threads=2,       
                use_mmap=False,    
                verbose=True
            )
            logger.info("✅ Brain Linked! System Online.")
        except Exception as e:
            logger.error(f"❌ Initialization failed: {e}")

app = FastAPI(title="ChatGPT Open-Source 1.0 API")

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

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

@app.get("/")
async def root():
    return {"status": "online", "message": "Connect to /chat"}

@app.post("/chat")
async def chat(request: Request):
    if llm is None:
        return JSONResponse({"response": "I'm still waking up..."}, status_code=503)
    
    try:
        data = await request.json()
        user_message = data.get("message")
        session_id = data.get("session_id", "default_user")
        
        if not user_message:
            return JSONResponse({"response": "No message received."}, status_code=400)

        # --- 4. Memory Logic ---
        if session_id not in sessions:
            sessions[session_id] = []
        
        history_str = ""
        for msg in sessions[session_id]:
            role = msg["role"]
            content = msg["content"]
            history_str += f"<|start|>{role}<|message|>{content}<|end|>\n"

        # --- 5. THE IRONCLAD PROMPT (Harmony Format) ---
        # We use 'developer' instructions because they override 'system' defaults.
        prompt = (
            f"<|start|>system<|message|>Knowledge cutoff: 2024-06. Current date: 2026-02-05. Reasoning: low.<|end|>\n"
            f"You are an independent AI assistant running on a high-performance cloud server. "
            f"{history_str}"
            f"<|start|>user<|message|>{user_message}<|end|>\n"
            f"<|start|>assistant<|channel|>final<|message|>"
        )
        
        output = llm(
            prompt,
            max_tokens=256,
            stop=["<|end|>", "<|return|>", "<|start|>", "user:", "assistant:", "Note:"],
            temperature=0.4 # Lower temperature = less hallucination/rambling
        )
        
        reply = output["choices"][0]["text"].strip()

        # Update Session Memory
        sessions[session_id].append({"role": "user", "content": user_message})
        sessions[session_id].append({"role": "assistant", "content": reply})
        
        if len(sessions[session_id]) > MAX_HISTORY:
            sessions[session_id] = sessions[session_id][-MAX_HISTORY:]

        return {"response": reply}

    except Exception as e:
        logger.error(f"❌ Error: {e}")
        return JSONResponse({"response": "error. Try shortening your message."}, status_code=500)

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)