Upload 2 files

app.py

@@ -0,0 +1,208 @@
+# -*- coding: utf-8 -*-
+# 財政部財政資訊中心 江信宗
+
+import gradio as gr
+import openai
+import time
+import re
+import os
+
+MODELS = [
+    "Meta-Llama-3.1-405B-Instruct",
+    "Meta-Llama-3.1-70B-Instruct",
+    "Meta-Llama-3.1-8B-Instruct"
+]
+
+API_BASE = "https://api.sambanova.ai/v1"
+
+def create_client(api_key=None):
+    """Creates an OpenAI client instance."""
+    if api_key:
+        openai.api_key = api_key
+    else:
+        openai.api_key = os.getenv("YOUR_API_TOKEN")
+    return openai.OpenAI(api_key=openai.api_key, base_url=API_BASE)
+
+def chat_with_ai(message, chat_history, system_prompt):
+    """Formats the chat history for the API call."""
+    # 初始化訊息列表，首先新增系統提示詞
+    messages = [{"role": "system", "content": system_prompt}]
+    # 遍歷聊天歷史，將使用者和AI機器人的對話新增到訊息列表中
+    for tup in chat_history:
+        # 獲取字典的第一個鍵（通常是使用者的訊息）
+        first_key = list(tup.keys())[0]
+        # 獲取字典的最後一個鍵（通常是AI機器人的回應）
+        last_key = list(tup.keys())[-1]
+        # 將使用者的訊息新增到messages列表中
+        messages.append({"role": "user", "content": tup[first_key]})
+        # 將AI機器人的回應新增到messages列表中
+        messages.append({"role": "assistant", "content": tup[last_key]})
+    # 新增當前使用者的訊息
+    messages.append({"role": "user", "content": message})
+    return messages
+
+def respond(message, chat_history, model, system_prompt, thinking_budget, api_key):
+    """Sends the message to the API and gets the response."""
+    # 建立OpenAI客戶端
+    client = create_client(api_key)
+    # 格式化聊天歷史
+    messages = chat_with_ai(message, chat_history, system_prompt.format(budget=thinking_budget))
+    # 記錄開始時間
+    start_time = time.time()
+    try:
+        # 呼叫API獲取回應
+        completion = client.chat.completions.create(model=model, messages=messages)
+        response = completion.choices[0].message.content
+        # 計算思考時間
+        thinking_time = time.time() - start_time
+        return response, thinking_time
+    except Exception as e:
+        # 捕獲並返回錯誤資訊
+        error_message = f"Error: {str(e)}"
+        return error_message, time.time() - start_time
+
+def parse_response(response):
+    """Parses the response from the API."""
+    # 使用正規表示式提取回答部分
+    answer_match = re.search(r'<answer>(.*?)</answer>', response, re.DOTALL)
+    # 使用正規表示式提取反思部分
+    reflection_match = re.search(r'<reflection>(.*?)</reflection>', response, re.DOTALL)
+    # 提取答案和反思內容，如果沒有匹配則設爲空字串
+    answer = answer_match.group(1).strip() if answer_match else ""
+    reflection = reflection_match.group(1).strip() if reflection_match else ""
+    # 提取所有步驟
+    steps = re.findall(r'<step>(.*?)</step>', response, re.DOTALL)
+    # 如果沒有提取到答案，則返回原始回應
+    if answer == "":
+        return response, "", ""
+    return answer, reflection, steps
+
+def generate(message, history, model, system_prompt, thinking_budget, api_key):
+    """Generates the chatbot response."""
+    # 獲取AI回應和思考時間
+    response, thinking_time = respond(message, history, model, system_prompt, thinking_budget, api_key)
+    # 如果回應是錯誤資訊，直接返回
+    if response.startswith("Error:"):
+        return history + [({"role": "system", "content": response},)], ""
+    # 解析AI的回應
+    answer, reflection, steps = parse_response(response)
+    # 初始化訊息列表
+    messages = []
+    # 新增使用者的輸入
+    messages.append({"role": "user", "content": f'<div style="text-align: left;">{message}</div>'})
+    # 格式化AI回應的步驟和反思
+    formatted_steps = [f"Step {i}：{step}" for i, step in enumerate(steps, 1)]
+    all_steps = "<br>".join(formatted_steps) + f"<br><br>Reflection：{reflection}"
+    # 新增AI的推理過程和思考時間
+    messages.append({
+        "role": "assistant", 
+        "content": f'<div style="text-align: left;">{all_steps}</div>', 
+        "metadata": {"title": f"推理過程時間： {thinking_time:.2f} 秒"}
+    })
+    # 新增AI的最終答案
+    messages.append({"role": "assistant", "content": f"<b>{answer}</b>"})
+    # 返回更新後的歷史記錄和空字串（用於清空輸入框）
+    return history + messages, ""
+
+DEFAULT_SYSTEM_PROMPT = """You are a helpful assistant in normal conversation.
+When given a problem to solve, you are an expert problem-solving assistant.
+Your task is to provide a detailed, step-by-step solution to a given question.
+Follow these instructions carefully:
+1. Read the given question carefully and reset counter between <count> and </count> to {budget}
+2. Generate a detailed, logical step-by-step solution.
+3. Enclose each step of your solution within <step> and </step> tags.
+4. You are allowed to use at most {budget} steps (starting budget),
+   keep track of it by counting down within tags <count> </count>,
+   STOP GENERATING MORE STEPS when hitting 0, you don't have to use all of them.
+5. Do a self-reflection when you are unsure about how to proceed,
+   based on the self-reflection and reward, decides whether you need to return
+   to the previous steps.
+6. After completing the solution steps, reorganize and synthesize the steps
+   into the final answer within <answer> and </answer> tags.
+7. Provide a critical, honest and subjective self-evaluation of your reasoning
+   process within <reflection> and </reflection> tags.
+8. Assign a quality score to your solution as a float between 0.0 (lowest
+   quality) and 1.0 (highest quality), enclosed in <reward> and </reward> tags.
+Example format:
+<count> [starting budget] </count>
+<step> [Content of step 1] </step>
+<count> [remaining budget] </count>
+<step> [Content of step 2] </step>
+<reflection> [Evaluation of the steps so far] </reflection>
+<reward> [Float between 0.0 and 1.0] </reward>
+<count> [remaining budget] </count>
+<step> [Content of step 3 or Content of some previous step] </step>
+<count> [remaining budget] </count>
+...
+<step>  [Content of final step] </step>
+<count> [remaining budget] </count>
+<answer> [Final Answer] </answer> (must give final answer in this format)
+<reflection> [Evaluation of the solution] </reflection>
+<reward> [Float between 0.0 and 1.0] </reward>
+"""
+
+custom_css = """
+.custom-button {
+    border-radius: 10px !important;
+}
+.natural-bg {
+    background-color: #e8f5e9 !important;
+    padding: 15px !important;
+    border-radius: 10px !important;
+}
+.pink-bg {
+    background-color: #ff4081 !important;
+    border-radius: 10px !important;
+}
+.pink-bg label, .pink-bg .label-wrap {
+    color: white !important;
+}
+.pink-bg textarea {
+    color: black !important;
+}
+.user-message .message.user {
+    background-color: #ffe0b2 !important;
+    border-radius: 10px !important;
+    padding: 10px !important;
+    margin: 0 !important;
+}
+.assistant-message .message.bot {
+    background-color: #e3f2fd !important;
+    border-radius: 5px !important;
+    padding: 10px !important;
+    margin: 0 !important;
+}
+"""
+
+with gr.Blocks(theme=gr.themes.Monochrome(), css=custom_css) as demo:
+    gr.Markdown("# Reasoning Chains using Llama-3.1-Instruct. Deployed by 江信宗")
+
+    with gr.Row(elem_classes="natural-bg"):
+        model = gr.Dropdown(choices=MODELS, label="選擇模型", value=MODELS[0])
+        thinking_budget = gr.Slider(minimum=1, maximum=100, value=20, step=1, label="思維規劃", info="模型所能進行的最大思考次數")
+        api_key = gr.Textbox(label="API Key", type="password", placeholder="API authentication key for large language models")
+
+    chatbot = gr.Chatbot(
+        label="Chat",
+        show_label=False,
+        show_share_button=False,
+        show_copy_button=True,
+        likeable=True,
+        layout="panel",
+        type="messages",
+        elem_classes=["user-message", "assistant-message"]
+    )
+    msg = gr.Textbox(label="在此輸入您的問題：", placeholder="輸入完成後直接按 Enter 開始執行......", elem_classes="pink-bg")
+    clear_button = gr.Button("清除聊天記錄", elem_classes="custom-button")
+    clear_button.click(lambda: ([], ""), inputs=None, outputs=[chatbot, msg])
+    # system_prompt = gr.Textbox(label="System Prompt", value=DEFAULT_SYSTEM_PROMPT, lines=15, interactive=True)
+    system_prompt = gr.Textbox(label="System Prompt", value=DEFAULT_SYSTEM_PROMPT, visible=False)
+    msg.submit(generate, inputs=[msg, chatbot, model, system_prompt, thinking_budget, api_key], outputs=[chatbot, msg])
+
+demo.load(js=None)
+
+if __name__ == "__main__":
+    if "SPACE_ID" in os.environ:
+        demo.launch()
+    else:
+        demo.launch(share=True, show_api=False)

requirements.txt

@@ -0,0 +1 @@
+openai