app.py

import os
import torch  # torch를 임포트
import faiss
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from sentence_transformers import SentenceTransformer
from datasets import load_dataset
import gradio as gr
from accelerate import Accelerator

# 환경 변수에서 Hugging Face API 키 로드
hf_api_key = os.getenv('HF_API_KEY')

# 모델 ID 및 토크나이저 설정
model_id = "mistralai/Mixtral-8x7B-Instruct-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id, token=hf_api_key)
accelerator = Accelerator()  # Accelerator 인스턴스 생성

# 모델 로딩
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    token=hf_api_key,
    torch_dtype=torch.bfloat16,  # torch를 사용해 데이터 타입 지정
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.bfloat16
    )
)
model = accelerator.prepare(model)  # 모델을 Accelerator에 준비시킴

# 데이터 로딩 및 faiss 인덱스 생성
ST = SentenceTransformer("mixedbread-ai/mxbai-embed-large-v1")
dataset = load_dataset("not-lain/wikipedia", revision="embedded")
data = dataset["train"]
data = data.add_faiss_index("embeddings")

# 검색 및 응답 생성 함수
def search(query: str, k: int = 3):
    embedded_query = ST.encode(query)
    scores, retrieved_examples = data.get_nearest_examples("embeddings", embedded_query, k=k)
    return scores, retrieved_examples

# 나머지 코드는 이전과 동일하게 유지


def format_prompt(prompt, retrieved_documents, k):
    PROMPT = f"Question:{prompt}\nContext:"
    for idx in range(k):
        PROMPT += f"{retrieved_documents['text'][idx]}\n"
    return PROMPT

def generate(formatted_prompt):
    formatted_prompt = formatted_prompt[:2000]  # GPU 메모리 제한을 고려
    messages = [{"role": "system", "content": "You are an assistant..."}, {"role": "user", "content": formatted_prompt}]
    input_ids = tokenizer(messages, return_tensors="pt", padding=True).input_ids.to(accelerator.device)
    outputs = model.generate(
        input_ids,
        max_new_tokens=1024,
        eos_token_id=tokenizer.eos_token_id,
        do_sample=True,
        temperature=0.6,
        top_p=0.9
    )
    response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
    return response

def rag_chatbot_interface(prompt: str, k: int = 2):
    scores, retrieved_documents = search(prompt, k)
    formatted_prompt = format_prompt(prompt, retrieved_documents, k)
    return generate(formatted_prompt)

SYS_PROMPT = "You are an assistant for answering questions. You are given the extracted parts of a long document and a question. Provide a conversational answer. If you don't know the answer, just say 'I do not know.' Don't make up an answer."

iface = gr.Interface(
    fn=rag_chatbot_interface,
    inputs=gr.inputs.Textbox(label="Enter your question"),
    outputs=gr.outputs.Textbox(label="Answer"),
    title="Retrieval-Augmented Generation Chatbot",
    description="This chatbot uses a retrieval-augmented generation approach to provide more accurate answers. It first searches for relevant documents and then generates a response based on the prompt and the retrieved documents."
)

iface.launch()