app.py

import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from sentence_transformers import SentenceTransformer
from datasets import load_dataset
import faiss
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
# 모델 ID
model_id = "microsoft/phi-2"

# 사용자 정의 코드를 신뢰하고 실행하도록 설정
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True  # 사용자 정의 코드 실행 허용
)
tokenizer = AutoTokenizer.from_pretrained(model_id, token=hf_api_key)
accelerator = Accelerator()

# 양자화 설정 없이 모델 로드 (문제 해결을 위한 임시 조치)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    token=hf_api_key,
    torch_dtype=torch.float32  # 기본 dtype 사용
)
model = accelerator.prepare(model)

# 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")

# 기타 함수 및 Gradio 인터페이스 구성은 이전과 동일


# Define functions for search, prompt formatting, and generation
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

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]  # Limit due to GPU memory constraints
    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
    )
    return tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)

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)

# Define system prompt for the chatbot
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="text",  # 텍스트 입력
    outputs="text",  # 텍스트 출력
    title="Retrieval-Augmented Generation Chatbot",
    description="This is a chatbot that 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()