app.py

import json
import gradio as gr
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity

import requests

def download_file(url, save_path):
    response = requests.get(url, stream=True)  # 스트리밍 모드로 받아서 메모리 절약
    response.raise_for_status()  # 오류 있으면 예외 발생
    
    with open(save_path, 'wb') as f:
        for chunk in response.iter_content(chunk_size=8192):
            if chunk:
                f.write(chunk)
    print(f"파일이 {save_path}에 저장되었어요!")

# 사용 예시
download_file('https://huggingface.co/datasets/Yuchan5386/AI-hub-SummaryData/resolve/main/g1.jsonl?download=true', 'data.jsonl')
file_path = 'data.jsonl'

# 1. 대화 데이터 불러오기 (모든 human-gpt 쌍 추출)
def load_conversations(file_path):
    conversations = []
    with open(file_path, 'r', encoding='utf-8') as f:
        for line in f:
            data = json.loads(line)
            conv = data.get('conversations', [])
            prev = None
            for c in conv:
                if c.get('from') == 'human':
                    prev = c['value']
                elif c.get('from') == 'gpt' and prev:
                    conversations.append((prev, c['value']))
                    prev = None
    return conversations

# 2. 한국어 불용어 설정 + TF-IDF 벡터라이저 준비
def build_vectorizer(questions):
    korean_stopwords = ['그리고', '그러나', '하지만', '또한', '이런', '저런', '그런', '있는', '없는', '것', '수', '때문에', '해서']
    vectorizer = TfidfVectorizer(stop_words=korean_stopwords)
    tfidf_matrix = vectorizer.fit_transform(questions)
    return vectorizer, tfidf_matrix

# 3. 유사도 기반 Top-k 응답 찾기
def find_top_responses(user_input, conversations, vectorizer, tfidf_matrix, top_k=3):
    user_vec = vectorizer.transform([user_input])
    sims = cosine_similarity(user_vec, tfidf_matrix)[0]
    top_indices = sims.argsort()[-top_k:][::-1]
    results = ""
    for i, idx in enumerate(top_indices, 1):
        answer = conversations[idx][1]
        score = sims[idx]
        results += f"{i}. {answer} (유사도: {score:.2f})\n\n"
    return results.strip()

# 초기화
convs = load_conversations(file_path)
questions = [q for q, _ in convs]
vectorizer, tfidf_matrix = build_vectorizer(questions)

# Gradio 인터페이스 함수
def chatbot_interface(user_input):
    if user_input.strip().lower() == "종료":
        return "ChatBot: 안녕~ 또 놀러와!"
    return find_top_responses(user_input, convs, vectorizer, tfidf_matrix)

# Gradio UI 구성
with gr.Blocks() as demo:
    gr.Markdown("##  ChatBot (TF-IDF 기반 Top-3 답변 추천)")
    with gr.Row():
        with gr.Column():
            input_box = gr.Textbox(label="질문을 입력해줘!", placeholder="예: 챗봇 어떻게 만들어요?")
            submit_btn = gr.Button("Toki에게 물어보기 🧠")
        with gr.Column():
            output_box = gr.Textbox(label="Toki의 추천 답변 Top-3", lines=10)

    submit_btn.click(fn=chatbot_interface, inputs=input_box, outputs=output_box)

demo.launch()