app.py

from fastapi import FastAPI, Header, HTTPException
import torch
import json
from transformers import AutoModelForCausalLM, AutoTokenizer
from duckduckgo_search import DDGS

app = FastAPI()

# පද්ධතියේ මතකය (Storage සඳහා)
LEARNING_FILE = "/data/elephant_learning_data.jsonl" # HF Storage path

# 18GB RAM එකට ගැලපෙන පරිදි Mistral 4-bit වලින් Load කිරීම
model_id = "mistralai/Mistral-7B-v0.3"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.bfloat16, 
    device_map="auto",
    load_in_4bit=True
)

# API Keys 50 (ELE-PRIME-001 to ELE-PRIME-050)
API_KEYS = {f"ELE-PRIME-{i:03d}": {"credits": 5000} for i in range(1, 51)}

@app.get("/")
def read_root():
    return {"message": "Elephant API Node 2026 is Online"}

@app.post("/v1/chat")
async def chat(message: dict, x_api_key: str = Header(None)):
    if x_api_key not in API_KEYS:
        raise HTTPException(status_code=403, detail="Invalid API Key")

    user_query = message.get("query", "")
    
    # Web Search for 2026 Live Data
    context = ""
    try:
        with DDGS() as ddgs:
            results = [r['body'] for r in ddgs.text(user_query, max_results=2)]
            context = "\n".join(results)
    except:
        context = "No live data available."

    # Response Generation
    input_text = f"Context: {context}\nUser: {user_query}\nAssistant:"
    inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=256)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True).split("Assistant:")[-1].strip()

    # Learning Loop: දත්ත පසුව Fine-tuning සඳහා Save කිරීම
    log_data = {"q": user_query, "a": response, "key": x_api_key}
    with open("learning_log.jsonl", "a") as f:
        f.write(json.dumps(log_data) + "\n")

    return {"reply": response, "status": "learned"}

main = app