app.py

import torch
import asyncio
import logging
import signal
import uvicorn
import os 

from fastapi import FastAPI, Request, HTTPException, status
from pydantic import BaseModel, Field
from langdetect import detect

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, GenerationConfig
from langchain.vectorstores import Qdrant
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.chains import RetrievalQA
from langchain.llms import HuggingFacePipeline
from qdrant_client import QdrantClient
from langchain.callbacks.base import BaseCallbackHandler
from huggingface_hub import hf_hub_download
from contextlib import asynccontextmanager

# Get environment variables
COLLECTION_NAME = "arabic_rag_collection"
QDRANT_URL = os.getenv("QDRANT_URL", "https://12efeef2-9f10-4402-9deb-f070977ddfc8.eu-central-1-0.aws.cloud.qdrant.io:6333")
QDRANT_API_KEY = os.getenv("QDRANT_API_KEY", "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJhY2Nlc3MiOiJtIn0.Jb39rYQW2rSE9RdXrjdzKY6T1RF44XjdQzCvzFkjat4")

# === LOGGING === #
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger(__name__)

# Load model and tokenizer
model_name = "FreedomIntelligence/Apollo-2B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token


# FastAPI setup
app = FastAPI(title="Apollo RAG Medical Chatbot")


# Generation settings
generation_config = GenerationConfig(
    max_new_tokens=150,
    temperature=0.2,
    top_k=20,
    do_sample=True,
    top_p=0.7,
    repetition_penalty=1.3,
)

# Text generation pipeline
llm_pipeline = pipeline(
    model=model,
    tokenizer=tokenizer,
    task="text-generation",
    generation_config=generation_config,
    device=model.device.index if model.device.type == "cuda" else -1
)

llm = HuggingFacePipeline(pipeline=llm_pipeline)

# Connect to Qdrant + embedding
embedding = HuggingFaceEmbeddings(model_name="Omartificial-Intelligence-Space/GATE-AraBert-v1")
qdrant_client = QdrantClient(url=QDRANT_URL, api_key=QDRANT_API_KEY)

vector_store = Qdrant(
    client=qdrant_client,
    collection_name=COLLECTION_NAME,
    embeddings=embedding
)

retriever = vector_store.as_retriever(search_kwargs={"k": 3})

# Set up RAG QA chain
qa_chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=retriever,
    chain_type="stuff"
)

class Query(BaseModel):
    question: str = Field(..., example="ما هي اسباب تساقط الشعر ؟", min_length=3)

class TimeoutCallback(BaseCallbackHandler):
    def __init__(self, timeout_seconds: int = 60):
        self.timeout_seconds = timeout_seconds
        self.start_time = None

    async def on_llm_start(self, *args, **kwargs):
        self.start_time = asyncio.get_event_loop().time()

    async def on_llm_new_token(self, *args, **kwargs):
        if asyncio.get_event_loop().time() - self.start_time > self.timeout_seconds:
            raise TimeoutError("LLM processing timeout")


# def generate_prompt(question: str) -> str:
#     lang = detect(question)
#     if lang == "ar":
#         return (
#             "أجب على السؤال الطبي التالي بلغة عربية فصحى، بإجابة دقيقة ومفصلة. إذا لم تجد معلومات كافية في السياق، استخدم معرفتك الطبية السابقة. \n"
#             "- عدم تكرار أي نقطة أو عبارة أو كلمة\n"
#             "- وضوح وسلاسة كل نقطة\n"
#             "- تجنب الحشو والعبارات الزائدة\n"
#             f"\nالسؤال: {question}\nالإجابة:"
#         )
#     else:
#         return (
#             "Answer the following medical question in clear English with a detailed, non-redundant response. "
#             "Do not repeat ideas, phrases, or restate the question in the answer. If the context lacks relevant "
#             "information, rely on your prior medical knowledge. If the answer involves multiple points, list them "
#             "in concise and distinct bullet points:\n"
#             f"Question: {question}\nAnswer:"
#         )

def generate_prompt(question):
    lang = detect(question)
    if lang == "ar":
        return f"""أجب على السؤال الطبي التالي بلغة عربية فصحى، بإجابة دقيقة ومفصلة. إذا لم تجد معلومات كافية في السياق، استخدم معرفتك الطبية السابقة. 
 وتأكد من ان:
- عدم تكرار أي نقطة أو عبارة أو كلمة
- وضوح وسلاسة كل نقطة
- تجنب الحشو والعبارات الزائدة-

السؤال: {question}
الإجابة:
"""
        
    else:
        return f"""Answer the following medical question in clear English with a detailed, non-redundant response. Do not repeat ideas, phrases, or restate the question in the answer. If the context lacks relevant information, rely on your prior medical knowledge. If the answer involves multiple points, list them in concise and distinct bullet points:
Question: {question}
Answer:"""

# === ROUTES === #
@app.get("/")
async def root():
    return {"message": "Medical QA API is running!"}

@app.post("/ask")
async def ask(query: Query):
    try:
        logger.debug(f"Received question: {query.question}")
        prompt = generate_prompt(query.question)
        timeout_callback = TimeoutCallback(timeout_seconds=60)

        # docs = retriever.get_relevant_documents(query.question)
        # if not docs:
        #     logger.warning("No documents retrieved from Qdrant for the question.")
        # else:
        #     logger.debug(f"Retrieved documents: {[doc.page_content for doc in docs[:1]]}")

        loop = asyncio.get_event_loop()
        
        response = await asyncio.wait_for(
            # qa_chain.run(prompt, callbacks=[timeout_callback]),
            loop.run_in_executor(None, qa_chain.run, prompt),
            timeout=360
        )

        if not response:
            raise ValueError("Empty answer returned from model")

        answer = response.split("Answer:")[-1].strip() if "Answer:" in response else response.split("الإجابة:")[-1].strip()
        
        return {
            "status": "success",
            "response": response,
            "answer": answer,
            "language": detect(query.question)
        }

    except TimeoutError as te:
        logger.error("Request timed out", exc_info=True)
        raise HTTPException(
            status_code=status.HTTP_504_GATEWAY_TIMEOUT,
            detail={"status": "error", "message": "Request timed out", "error": str(te)}
        )

    except Exception as e:
        logger.error(f"Unexpected error: {e}", exc_info=True)
        raise HTTPException(
            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
            detail={"status": "error", "message": "Internal server error", "error": str(e)}
        )

@app.post("/chat")
def chat(query: Query):

    prompt = generate_prompt(query.question)

    response = qa_chain.run(prompt)

    answer = response.split("Answer:")[-1].strip() if "Answer:" in response else response.split("الإجابة:")[-1].strip()
    

    return {
        "response": response,
        "answer": answer
    }

    

# === ENTRYPOINT === #
if __name__ == "__main__":
    def handle_exit(signum, frame):
        print("Shutting down gracefully...")
        exit(0)

    signal.signal(signal.SIGINT, handle_exit)
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)



# from langdetect import detect
# from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline, GenerationConfig
# import torch
# import logging
# from fastapi import FastAPI, Request, HTTPException, status
# from pydantic import BaseModel, Field
# import time
# import asyncio
# from concurrent.futures import ThreadPoolExecutor
# from fastapi.middleware.cors import CORSMiddleware

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

# # Load model and tokenizer
# model_name = "FreedomIntelligence/Apollo-7B"
# # model_name = "emilyalsentzer/Bio_ClinicalBERT"
# # model_name = "FreedomIntelligence/Apollo-2B"

# tokenizer = AutoTokenizer.from_pretrained(model_name)
# model = AutoModelForCausalLM.from_pretrained(model_name)

# tokenizer.pad_token = tokenizer.eos_token

# app = FastAPI(title="Apollo RAG Medical Chatbot")

# # Add this after creating the `app`
# app.add_middleware(
#     CORSMiddleware,
#     allow_origins=["*"],  # Allow all origins
#     allow_credentials=True,
#     allow_methods=["*"],
#     allow_headers=["*"],
# )



# generation_config = GenerationConfig(
#     max_new_tokens=150,
#     temperature=0.2,
#     top_k=20,
#     do_sample=True,
#     top_p=0.7,
#     repetition_penalty=1.3,
# )

# # Create generation pipeline
# pipe = TextGenerationPipeline(
#     model=model,
#     tokenizer=tokenizer,
#     device=model.device.index if torch.cuda.is_available() else "cpu"
# )

# # Prompt formatter based on language
# def generate_prompt(message):
#     lang = detect(message)
#     if lang == "ar":
#         return f"""أجب على السؤال الطبي التالي بلغة عربية فصحى، بإجابة دقيقة ومفصلة. إذا لم تجد معلومات كافية في السياق، استخدم معرفتك الطبية السابقة.
#  وتأكد من ان:
# - عدم تكرار أي نقطة أو عبارة أو كلمة
# - وضوح وسلاسة كل نقطة
# - تجنب الحشو والعبارات الزائدة
# السؤال: {message}
# الإجابة:"""
#     else:
#         return f"""Answer the following medical question in clear English with a detailed, non-redundant response. Do not repeat ideas or restate the question. If information is missing, rely on your prior medical knowledge:
# Question: {message}
# Answer:"""

# # Chat function
# # @app.post("/ask")
# # def chat_fn(message):
# #     prompt = generate_prompt(message)
# #     response = pipe(prompt,
# #                     max_new_tokens=512,
# #                     temperature=0.7,
# #                     do_sample = True,
# #                     top_p=0.9)[0]['generated_text']
# #     answer = response.split("Answer:")[-1].strip() if "Answer:" in response else response.split("الإجابة:")[-1].strip()
# #     return {"Answer": answer}

# executor = ThreadPoolExecutor()

# # Define request model
# class Query(BaseModel):
#     message: str

# @app.get("/")
# def read_root():
#     return {"message": "Apollo Medical Chatbot API is running"}


# # @app.post("/ask")
# # async def chat_fn(query: Query):
    
# #     message = query.message
# #     logger.info(f"Received message: {message}")
    
# #     prompt = generate_prompt(message)

# #     # Run blocking inference in thread
# #     loop = asyncio.get_event_loop()
# #     response = await loop.run_in_executor(executor,
# #                                           lambda: pipe(prompt, max_new_tokens=512, temperature=0.7, do_sample=True, top_p=0.9)[0]['generated_text'])

# #     # Parse answer
# #     answer = response.split("Answer:")[-1].strip() if "Answer:" in response else response.split("الإجابة:")[-1].strip()
# #     return {"Answer": answer}

# @app.post("/ask")
# async def chat_fn(query: Query):
#     message = query.message
#     logger.info(f"Received message: {message}")
    
#     prompt = generate_prompt(message)

#     try:
#         start_time = time.time()
        
#         loop = asyncio.get_event_loop()
#         response = await loop.run_in_executor(
#             executor,
#             lambda: pipe(prompt, max_new_tokens=150, temperature=0.7, do_sample=True, top_p=0.9)[0]['generated_text']
#         )
        
#         duration = time.time() - start_time
#         logger.info(f"Model inference completed in {duration:.2f} seconds")

#         logger.info(f"Generated answer: {answer}")
#         return {"Answer": answer}

#     except Exception as e:
#         logger.error(f"Inference failed: {str(e)}")
#         raise HTTPException(status_code=500, detail="Model inference TimeOut failed.")