app.py

import os
import gradio as gr
from kiwipiepy import Kiwi

from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough, RunnableLambda
from langchain_community.document_transformers import LongContextReorder

from libs.config import STREAMING
from libs.embeddings import get_embeddings
from libs.retrievers import load_retrievers
from libs.llm import get_llm
from libs.prompt import get_prompt

os.environ["TOKENIZERS_PARALLELISM"] = "false"


def kiwi_tokenize(text):
    kiwi = Kiwi()
    return [token.form for token in kiwi.tokenize(text)]


embeddings = get_embeddings()
retriever = load_retrievers(embeddings)


def kiwi_tokenize(text):
    kiwi = Kiwi()
    return [token.form for token in kiwi.tokenize(text)]


embeddings = get_embeddings()
retriever = load_retrievers(embeddings)


def create_rag_chain(chat_history, model):
    llm = get_llm(streaming=STREAMING).with_config(configurable={"llm": model})
    prompt = get_prompt(chat_history)

    return (
        {
            "context": retriever
            | RunnableLambda(LongContextReorder().transform_documents),
            "question": RunnablePassthrough(),
        }
        | prompt
        | llm
        | StrOutputParser()
    )


def respond_stream(message, history, model):
    rag_chain = create_rag_chain(history, model)
    for chunk in rag_chain.stream(message):
        yield chunk


def respond(message, history, model):
    rag_chain = create_rag_chain(history, model)
    return rag_chain.invoke(message)


# 사용 가능한 모델 목록 (key: 모델 식별자, value: 사용자에게 표시할 레이블)
AVAILABLE_MODELS = {
    "gpt_3_5_turbo": "GPT-3.5 Turbo",
    "gpt_4o": "GPT-4o",
    "claude_3_5_sonnet": "Claude 3.5 Sonnet",
    "gemini_1_5_flash": "Gemini 1.5 Flash",
    "llama3_70b": "Llama3 70b",
}


def get_model_key(label):
    return next(key for key, value in AVAILABLE_MODELS.items() if value == label)


def chat_function(message, history, model_label):
    model_key = get_model_key(model_label)
    if STREAMING:
        response = ""
        for chunk in respond_stream(message, history, model_key):
            response += chunk
            yield response
    else:
        response = respond(message, history, model_key)
        yield response


with gr.Blocks() as demo:
    gr.Markdown("# 대법원 판례 상담 도우미")
    gr.Markdown(
        "안녕하세요! 대법원 판례에 관한 질문에 답변해드리는 AI 상담 도우미입니다. 판례 검색, 해석, 적용 등에 대해 궁금하신 점이 있으면 언제든 물어보세요."
    )

    model_dropdown = gr.Dropdown(
        choices=list(AVAILABLE_MODELS.values()),
        label="모델 선택",
        value=list(AVAILABLE_MODELS.values())[1],
    )

    chatbot = gr.ChatInterface(
        fn=chat_function,
        additional_inputs=[model_dropdown],
    )

if __name__ == "__main__":
    demo.launch()