app.py

import gradio as gr
from huggingface_hub import InferenceClient
import openai
import anthropic
import os
from typing import Optional

#############################
# [기본코드] - 수정/삭제 불가
#############################

# Cohere Command R+ 모델 ID 정의
COHERE_MODEL = "CohereForAI/c4ai-command-r-plus-08-2024"

def get_client(model_name):
    """
    모델 이름에 맞춰 InferenceClient 생성.
    토큰은 환경 변수에서 가져옴.
    """
    hf_token = os.getenv("HF_TOKEN")
    if not hf_token:
        raise ValueError("HuggingFace API 토큰이 필요합니다.")

    if model_name == "Cohere Command R+":
        model_id = COHERE_MODEL
    else:
        raise ValueError("유효하지 않은 모델 이름입니다.")
    return InferenceClient(model_id, token=hf_token)

def respond_cohere_qna(
    question: str,
    system_message: str,
    max_tokens: int,
    temperature: float,
    top_p: float
):
    """
    Cohere Command R+ 모델을 이용해 한 번의 질문(question)에 대한 답변을 반환하는 함수.
    """
    model_name = "Cohere Command R+"
    try:
        client = get_client(model_name)
    except ValueError as e:
        return f"오류: {str(e)}"

    messages = [
        {"role": "system", "content": system_message},
        {"role": "user", "content": question}
    ]

    try:
        response_full = client.chat_completion(
            messages,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
        )
        assistant_message = response_full.choices[0].message.content
        return assistant_message
    except Exception as e:
        return f"오류가 발생했습니다: {str(e)}"

def respond_chatgpt_qna(
    question: str,
    system_message: str,
    max_tokens: int,
    temperature: float,
    top_p: float
):
    """
    ChatGPT(OpenAI) 모델을 이용해 한 번의 질문(question)에 대한 답변을 반환하는 함수.
    """
    openai_token = os.getenv("OPENAI_TOKEN")
    if not openai_token:
        return "OpenAI API 토큰이 필요합니다."

    openai.api_key = openai_token

    messages = [
        {"role": "system", "content": system_message},
        {"role": "user", "content": question}
    ]

    try:
        response = openai.ChatCompletion.create(
            model="gpt-4o-mini",
            messages=messages,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
        )
        assistant_message = response.choices[0].message['content']
        return assistant_message
    except Exception as e:
        return f"오류가 발생했습니다: {str(e)}"

def respond_deepseek_qna(
    question: str,
    system_message: str,
    max_tokens: int,
    temperature: float,
    top_p: float,
    model_name: str  # 모델 이름 추가
):
    """
    DeepSeek 모델을 이용해 한 번의 질문(question)에 대한 답변을 반환하는 함수.
    """
    deepseek_token = os.getenv("DEEPSEEK_TOKEN")
    if not deepseek_token:
        return "DeepSeek API 토큰이 필요합니다."

    openai.api_key = deepseek_token
    openai.api_base = "https://api.deepseek.com/v1"

    messages = [
        {"role": "system", "content": system_message},
        {"role": "user", "content": question}
    ]

    try:
        response = openai.ChatCompletion.create(
            model=model_name,  # 선택된 모델 사용
            messages=messages,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
        )
        assistant_message = response.choices[0].message['content']
        return assistant_message
    except Exception as e:
        return f"오류가 발생했습니다: {str(e)}"

def respond_claude_qna(
    question: str,
    system_message: str,
    max_tokens: int,
    temperature: float,
    top_p: float,
    model_name: str  # 모델 이름 파라미터 추가
) -> str:
    """
    Claude API를 사용한 개선된 응답 생성 함수
    """
    claude_api_key = os.getenv("CLAUDE_TOKEN")
    if not claude_api_key:
        return "Claude API 토큰이 필요합니다."

    try:
        client = anthropic.Anthropic(api_key=claude_api_key)
        
        message = client.messages.create(
            model=model_name,
            max_tokens=max_tokens,
            temperature=temperature,
            system=system_message,
            messages=[
                {"role": "user", "content": question}
            ]
        )
        
        return message.content[0].text
            
    except anthropic.APIError as ae:
        return f"Claude API 오류: {str(ae)}"
    except anthropic.RateLimitError:
        return "요청 한도를 초과했습니다. 잠시 후 다시 시도해주세요."
    except Exception as e:
        return f"예상치 못한 오류가 발생했습니다: {str(e)}"

def respond_o1mini_qna(
    question: str,
    system_message: str,
    max_tokens: int,
    temperature: float
):
    """
    o1-mini 모델을 이용해 한 번의 질문(question)에 대한 답변을 반환하는 함수.
    o1-mini에서는 'system' 역할 메시지를 지원하지 않으므로,
    system_message와 question을 하나의 'user' 메시지로 합쳐 전달합니다.
    또한, o1-mini에서는 'max_tokens' 대신 'max_completion_tokens' 파라미터를 사용하며,
    온도(temperature)는 고정값 1만 지원합니다.
    """
    openai_token = os.getenv("OPENAI_TOKEN")
    if not openai_token:
        return "OpenAI API 토큰이 필요합니다."

    openai.api_key = openai_token

    combined_message = f"{system_message}\n\n{question}"
    messages = [
        {"role": "user", "content": combined_message}
    ]

    try:
        response = openai.ChatCompletion.create(
            model="o1-mini",
            messages=messages,
            max_completion_tokens=max_tokens,
            temperature=1,  # 고정된 값 1 사용
        )
        assistant_message = response.choices[0].message['content']
        return assistant_message
    except Exception as e:
        return f"오류가 발생했습니다: {str(e)}"

def respond_gemini_qna(
    question: str,
    system_message: str,
    max_tokens: int,
    temperature: float,
    top_p: float,
    model_id: str
):
    """
    Gemini 모델(예: gemini-2.0-flash, gemini-2.0-flash-lite-preview-02-05)을 이용해
    한 번의 질문(question)에 대한 답변을 반환하는 함수.
    system_message와 question을 하나의 프롬프트로 합쳐서 사용합니다.
    """
    hf_token = os.getenv("HF_TOKEN")
    if not hf_token:
        return "HuggingFace API 토큰이 필요합니다."
    
    client = InferenceClient(model_id, token=hf_token)
    prompt = f"{system_message}\n\n{question}"
    
    try:
        response = client.text_generation(
            prompt,
            max_new_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p
        )
        return response["generated_text"]
    except Exception as e:
        return f"오류가 발생했습니다: {str(e)}"

#############################
# [기본코드] UI 부분 - 수정/삭제 불가 (탭 순서: OpenAI, Gemini, Claude, DeepSeek, Cohere Command R+)
#############################

with gr.Blocks() as demo:
    gr.Markdown("# LLM 플레이그라운드")

    #################
    # OpenAI 탭 (gpt-4o-mini / o1-mini 통합)
    #################
    with gr.Tab("OpenAI"):
        openai_model_radio = gr.Radio(
            choices=["gpt-4o-mini", "o1-mini"],
            label="모델 선택",
            value="gpt-4o-mini"
        )
        
        with gr.Column(visible=True) as chatgpt_ui:
            chatgpt_input1_o = gr.Textbox(label="입력1", lines=1)
            chatgpt_input2_o = gr.Textbox(label="입력2", lines=1)
            chatgpt_input3_o = gr.Textbox(label="입력3", lines=1)
            chatgpt_input4_o = gr.Textbox(label="입력4", lines=1)
            chatgpt_input5_o = gr.Textbox(label="입력5", lines=1)
            chatgpt_answer_output_o = gr.Textbox(label="결과", lines=5, interactive=False)
            with gr.Accordion("고급 설정 (gpt-4o-mini)", open=False):
                chatgpt_system_message_o = gr.Textbox(
                    value="""반드시 한글로 답변할 것.
너는 ChatGPT, OpenAI에서 개발한 언어 모델이다.
내가 요구하는 것을 최대한 자세하고 정확하게 답변하라.
""",
                    label="System Message",
                    lines=3
                )
                chatgpt_max_tokens_o = gr.Slider(minimum=100, maximum=4000, value=2000, step=100, label="Max Tokens")
                chatgpt_temperature_o = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.05, label="Temperature")
                chatgpt_top_p_o = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-P")
            chatgpt_submit_button_o = gr.Button("전송")
            
            def merge_and_call_chatgpt_o(i1, i2, i3, i4, i5, sys_msg, mt, temp, top_p_):
                question = " ".join([i1, i2, i3, i4, i5])
                return respond_chatgpt_qna(
                    question=question,
                    system_message=sys_msg,
                    max_tokens=mt,
                    temperature=temp,
                    top_p=top_p_
                )
            chatgpt_submit_button_o.click(
                fn=merge_and_call_chatgpt_o,
                inputs=[
                    chatgpt_input1_o, chatgpt_input2_o, chatgpt_input3_o, chatgpt_input4_o, chatgpt_input5_o,
                    chatgpt_system_message_o,
                    chatgpt_max_tokens_o,
                    chatgpt_temperature_o,
                    chatgpt_top_p_o
                ],
                outputs=chatgpt_answer_output_o
            )
        
        with gr.Column(visible=False) as o1mini_ui:
            o1mini_input1_o = gr.Textbox(label="입력1", lines=1)
            o1mini_input2_o = gr.Textbox(label="입력2", lines=1)
            o1mini_input3_o = gr.Textbox(label="입력3", lines=1)
            o1mini_input4_o = gr.Textbox(label="입력4", lines=1)
            o1mini_input5_o = gr.Textbox(label="입력5", lines=1)
            o1mini_answer_output_o = gr.Textbox(label="결과", lines=5, interactive=False)
            with gr.Accordion("고급 설정 (o1-mini)", open=False):
                o1mini_system_message_o = gr.Textbox(
                    value="""반드시 한글로 답변할 것.
너는 o1-mini, OpenAI에서 개발한 경량 언어 모델이다.
내가 요구하는 것을 최대한 자세하고 정확하게 답변하라.
""",
                    label="System Message",
                    lines=3
                )
                o1mini_max_tokens_o = gr.Slider(minimum=100, maximum=4000, value=2000, step=100, label="Max Tokens")
                o1mini_temperature_o = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.05, label="Temperature")
            o1mini_submit_button_o = gr.Button("전송")
            
            def merge_and_call_o1mini_o(i1, i2, i3, i4, i5, sys_msg, mt, temp):
                question = " ".join([i1, i2, i3, i4, i5])
                return respond_o1mini_qna(
                    question=question,
                    system_message=sys_msg,
                    max_tokens=mt,
                    temperature=temp
                )
            o1mini_submit_button_o.click(
                fn=merge_and_call_o1mini_o,
                inputs=[
                    o1mini_input1_o, o1mini_input2_o, o1mini_input3_o, o1mini_input4_o, o1mini_input5_o,
                    o1mini_system_message_o,
                    o1mini_max_tokens_o,
                    o1mini_temperature_o
                ],
                outputs=o1mini_answer_output_o
            )
        
        def update_openai_ui(model_choice):
            if model_choice == "gpt-4o-mini":
                return gr.update(visible=True), gr.update(visible=False)
            else:
                return gr.update(visible=False), gr.update(visible=True)
        
        openai_model_radio.change(
            fn=update_openai_ui,
            inputs=openai_model_radio,
            outputs=[chatgpt_ui, o1mini_ui]
        )

    #################
    # Gemini 탭
    #################
    with gr.Tab("Gemini"):
        gemini_model_radio = gr.Radio(
            choices=["gemini-2.0-flash", "gemini-2.0-flash-lite-preview-02-05"],
            label="모델 선택",
            value="gemini-2.0-flash"
        )
        gemini_input1 = gr.Textbox(label="입력1", lines=1)
        gemini_input2 = gr.Textbox(label="입력2", lines=1)
        gemini_input3 = gr.Textbox(label="입력3", lines=1)
        gemini_input4 = gr.Textbox(label="입력4", lines=1)
        gemini_input5 = gr.Textbox(label="입력5", lines=1)
        gemini_answer_output = gr.Textbox(label="결과", lines=5, interactive=False)
        with gr.Accordion("고급 설정 (Gemini)", open=False):
            gemini_system_message = gr.Textbox(
                value="""반드시 한글로 답변할 것.
너는 Gemini 모델이다.
내가 요구하는 것을 최대한 자세하고 정확하게 답변하라.
""",
                label="System Message",
                lines=3
            )
            gemini_max_tokens = gr.Slider(minimum=100, maximum=4000, value=2000, step=100, label="Max Tokens")
            gemini_temperature = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.05, label="Temperature")
            gemini_top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-P")
        gemini_submit_button = gr.Button("전송")
        
        def merge_and_call_gemini(i1, i2, i3, i4, i5, sys_msg, mt, temp, top_p_, model_radio):
            question = " ".join([i1, i2, i3, i4, i5])
            return respond_gemini_qna(
                question=question,
                system_message=sys_msg,
                max_tokens=mt,
                temperature=temp,
                top_p=top_p_,
                model_id=model_radio
            )
        gemini_submit_button.click(
            fn=merge_and_call_gemini,
            inputs=[
                gemini_input1, gemini_input2, gemini_input3, gemini_input4, gemini_input5,
                gemini_system_message,
                gemini_max_tokens,
                gemini_temperature,
                gemini_top_p,
                gemini_model_radio
            ],
            outputs=gemini_answer_output
        )

    #################
    # Claude 탭
    #################
    with gr.Tab("Claude"):
        claude_model_radio = gr.Radio(
            choices=[
                "claude-3-haiku-20240307",
                "claude-3-5-haiku-20241022",
                "claude-3-5-sonnet-20241022"
            ],
            label="모델 선택",
            value="claude-3-5-sonnet-20241022"
        )
    
        claude_input1 = gr.Textbox(label="입력1", lines=1)
        claude_input2 = gr.Textbox(label="입력2", lines=1)
        claude_input3 = gr.Textbox(label="입력3", lines=1)
        claude_input4 = gr.Textbox(label="입력4", lines=1)
        claude_input5 = gr.Textbox(label="입력5", lines=1)

        claude_answer_output = gr.Textbox(label="결과", interactive=False, lines=5)

        with gr.Accordion("고급 설정 (Claude)", open=False):
            claude_system_message = gr.Textbox(
                label="System Message",
                value="""반드시 한글로 답변할 것.
너는 Anthropic에서 개발한 클로드이다.
최대한 정확하고 친절하게 답변하라.""",
                lines=3
            )
            claude_max_tokens = gr.Slider(minimum=100, maximum=4000, value=2000, step=100, label="Max Tokens")
            claude_temperature = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.05, label="Temperature")
            claude_top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p")
        claude_submit_button = gr.Button("전송")

        def merge_and_call_claude(i1, i2, i3, i4, i5, sys_msg, mt, temp, top_p_, model_radio):
            question = " ".join([i1, i2, i3, i4, i5])
            return respond_claude_qna(
                question=question,
                system_message=sys_msg,
                max_tokens=mt,
                temperature=temp,
                top_p=top_p_,
                model_name=model_radio
            )
        claude_submit_button.click(
            fn=merge_and_call_claude,
            inputs=[
                claude_input1, claude_input2, claude_input3, claude_input4, claude_input5,
                claude_system_message,
                claude_max_tokens,
                claude_temperature,
                claude_top_p,
                claude_model_radio
            ],
            outputs=claude_answer_output
        )

    #################
    # DeepSeek 탭
    #################
    with gr.Tab("DeepSeek"):
        deepseek_model_radio = gr.Radio(
            choices=["V3 (deepseek-chat)", "R1 (deepseek-reasoner)"],
            label="모델 선택",
            value="V3 (deepseek-chat)"
        )

        deepseek_input1 = gr.Textbox(label="입력1", lines=1)
        deepseek_input2 = gr.Textbox(label="입력2", lines=1)
        deepseek_input3 = gr.Textbox(label="입력3", lines=1)
        deepseek_input4 = gr.Textbox(label="입력4", lines=1)
        deepseek_input5 = gr.Textbox(label="입력5", lines=1)

        deepseek_answer_output = gr.Textbox(label="결과", lines=5, interactive=False)

        with gr.Accordion("고급 설정 (DeepSeek)", open=False):
            deepseek_system_message = gr.Textbox(
                value="""반드시 한글로 답변할 것.
너는 DeepSeek-V3, 최고의 언어 모델이다.
내가 요구하는 것을 최대한 자세하고 정확하게 답변하라.
""",
                label="System Message",
                lines=3
            )
            deepseek_max_tokens = gr.Slider(minimum=100, maximum=4000, value=2000, step=100, label="Max Tokens")
            deepseek_temperature = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.05, label="Temperature")
            deepseek_top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-P")
        deepseek_submit_button = gr.Button("전송")

        def merge_and_call_deepseek(i1, i2, i3, i4, i5, sys_msg, mt, temp, top_p_, model_radio):
            if model_radio == "V3 (deepseek-chat)":
                model_name = "deepseek-chat"
            else:
                model_name = "deepseek-reasoner"
            question = " ".join([i1, i2, i3, i4, i5])
            return respond_deepseek_qna(
                question=question,
                system_message=sys_msg,
                max_tokens=mt,
                temperature=temp,
                top_p=top_p_,
                model_name=model_name
            )
        deepseek_submit_button.click(
            fn=merge_and_call_deepseek,
            inputs=[
                deepseek_input1, deepseek_input2, deepseek_input3, deepseek_input4, deepseek_input5,
                deepseek_system_message,
                deepseek_max_tokens,
                deepseek_temperature,
                deepseek_top_p,
                deepseek_model_radio
            ],
            outputs=deepseek_answer_output
        )

    #################
    # Cohere Command R+ 탭
    #################
    with gr.Tab("Cohere Command R+"):
        cohere_input1 = gr.Textbox(label="입력1", lines=1)
        cohere_input2 = gr.Textbox(label="입력2", lines=1)
        cohere_input3 = gr.Textbox(label="입력3", lines=1)
        cohere_input4 = gr.Textbox(label="입력4", lines=1)
        cohere_input5 = gr.Textbox(label="입력5", lines=1)

        cohere_answer_output = gr.Textbox(label="결과", lines=5, interactive=False)

        with gr.Accordion("고급 설정 (Cohere)", open=False):
            cohere_system_message = gr.Textbox(
                value="""반드시 한글로 답변할 것.
너는 최고의 비서이다.
내가 요구하는것들을 최대한 자세하고 정확하게 답변하라.
""",
                label="System Message",
                lines=3
            )
            cohere_max_tokens = gr.Slider(minimum=100, maximum=10000, value=4000, step=100, label="Max Tokens")
            cohere_temperature = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.1, label="Temperature")
            cohere_top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-P")
        cohere_submit_button = gr.Button("전송")

        def merge_and_call_cohere(i1, i2, i3, i4, i5, sys_msg, mt, temp, top_p_):
            question = " ".join([i1, i2, i3, i4, i5])
            return respond_cohere_qna(
                question=question,
                system_message=sys_msg,
                max_tokens=mt,
                temperature=temp,
                top_p=top_p_
            )
        cohere_submit_button.click(
            fn=merge_and_call_cohere,
            inputs=[
                cohere_input1, cohere_input2, cohere_input3, cohere_input4, cohere_input5,
                cohere_system_message,
                cohere_max_tokens,
                cohere_temperature,
                cohere_top_p
            ],
            outputs=cohere_answer_output
        )

#############################
# 메인 실행부
#############################
if __name__ == "__main__":
    demo.launch()