Spaces:

My-Project-Team
/

MachineVision_DeepLearning_FinalProject

Sleeping

MachineVision_DeepLearning_FinalProject / app.py

Update app.py

a0f68f2 verified 11 months ago

12.4 kB

import gradio as gr
from openai import OpenAI
from huggingface_hub import InferenceClient
import os
import time
import requests
import urllib.parse
import pandas as pd
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from PIL import Image
from dotenv import load_dotenv

# --- 匯入組員的模組 ---
import emotion 
# ---------------------------

load_dotenv()

# ==========================================
# 0. 環境變數 & 1. 系統初始化 (保持不變)
# ==========================================
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
HF_TOKEN = os.getenv("HF_TOKEN")

global_df = None
global_mood_df = None
global_retriever = None
rag_initialized = False

def init_rag_system():
    global global_df, global_mood_df, global_retriever, rag_initialized
    if rag_initialized: return
    try:
        global_df = pd.read_csv('restaurants.csv')
        global_df['RAG_Content'] = global_df['RAG_Content'].fillna("")
        global_df['Category'] = global_df['Category'].fillna("其他")
        global_mood_df = pd.read_csv('mood_food_guide.csv')
    except Exception: pass
    if os.path.exists("faiss_index"):
        try:
            embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
            vectorstore = FAISS.load_local("faiss_index", embeddings, allow_dangerous_deserialization=True)
            global_retriever = vectorstore.as_retriever(search_kwargs={"k": 3})
        except Exception: pass
    rag_initialized = True

# ==========================================
# 2. 核心功能 (修改：回傳分開的除錯資訊)
# ==========================================
def get_restaurant_data(mood_score_str, food_choice):
    init_rag_system()
    if global_df is None or global_df.empty: return None, True, "資料庫未載入", "無建議"
    try: score = int(str(mood_score_str).split(' ')[0])
    except: score = 3
    mood_info = global_mood_df[global_mood_df['分數'] == score]
    if not mood_info.empty:
        rec_categories = mood_info.iloc[0]['推薦料理類別']
        mood_reason = mood_info.iloc[0]['原因']
    else:
        rec_categories = ""
        mood_reason = "隨意探索"
    candidates = global_df.copy()
    if rec_categories:
        candidates = candidates[candidates['Category'].apply(lambda x: str(x) in str(rec_categories) or str(rec_categories) in str(x))]
    food_keyword = "飯" if food_choice == "吃飯" else "麵" if food_choice == "吃麵" else ""
    if food_keyword:
        candidates = candidates[candidates['Name'].str.contains(food_keyword, case=False, na=False) | candidates['RAG_Content'].str.contains(food_keyword, case=False, na=False)]
    if candidates.empty:
        result = global_df.sample(1).iloc[0]; is_random = True
    else:
        result = candidates.sample(1).iloc[0]; is_random = False
    return result, is_random, rec_categories, mood_reason

def generate_content_with_groq(restaurant_name, restaurant_detail, user_diary, mood_score, mood_guide_reason, debug_mode=False):
    if not GROQ_API_KEY: return "⚠️ 請設定 GROQ_API_KEY", ""
    
    client = OpenAI(api_key=GROQ_API_KEY, base_url="https://api.groq.com/openai/v1")
    
    system_prompt = "你是一個幽默、懂吃且善解人意的 AI 朋友..."
    user_msg = f"【狀態】心情分數：{mood_score}\n【日記】{user_diary}\n【建議原因】{mood_guide_reason}\n【餐廳】{restaurant_name}\n資料：{restaurant_detail}"
    
    # 準備除錯資訊
    debug_log = ""
    if debug_mode:
        debug_log = f"""
### 🔧 Groq Prompt Debug
**System Prompt:**
{system_prompt}
**User Message:**
{user_msg}
"""

    try:
        response = client.chat.completions.create(model="llama-3.3-70b-versatile", messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": user_msg}])
        content = response.choices[0].message.content
        return content, debug_log # 回傳 (內容, 除錯資訊)
    except Exception as e: 
        return f"Groq Error: {str(e)}", debug_log

def generate_image_huggingface(prompt):
    if not HF_TOKEN: return None
    try:
        hf_client = InferenceClient(token=HF_TOKEN)
        return hf_client.text_to_image(prompt=prompt, model="stabilityai/stable-diffusion-xl-base-1.0")
    except: return None

def mood_agent_logic(score_input, food_input, diary_input, debug_mode):
    restaurant, is_random, rec_categories, mood_reason = get_restaurant_data(score_input, food_input)
    if restaurant is None: yield "資料庫讀取錯誤", None, "", gr.update(); return
    
    name = restaurant['Name']; address = restaurant['Address']; url = restaurant['URL']; img_prompt = restaurant.get('Visual_prompt')
    
    rag_info = str(restaurant.get('RAG_Content', ''))
    if global_retriever:
        docs = global_retriever.invoke(name)
        if docs: rag_info = "\n".join([d.page_content for d in docs])
        
    # 呼叫 Groq 生成內容與取得除錯資訊
    ai_text, groq_debug_log = generate_content_with_groq(name, rag_info, diary_input, score_input, mood_reason, debug_mode)
    
    # 準備完整的除錯資訊 (包含圖片 Prompt)
    full_debug_log = ""
    if debug_mode:
        img_debug_log = f"""
### 🎨 Image Prompt Debug
**Visual Prompt:**
{img_prompt}
"""
        full_debug_log = groq_debug_log + "\n\n" + img_debug_log

    # 設定除錯視窗的狀態 (有勾選才顯示)
    debug_output_update = gr.update(value=full_debug_log, visible=debug_mode)

    final_response = f"### 🍽️ 推薦：{name}\n\n{ai_text}"
    map_html = f'<div style="text-align:center"><a href="{url}" target="_blank" style="background:#4CAF50;color:white;padding:8px 16px;border-radius:20px;text-decoration:none">🗺️ Google Map 導航</a></div>'
    
    # 先回傳文字與除錯資訊
    yield final_response, None, map_html, debug_output_update
    
    # 再回傳圖片
    image_output = generate_image_huggingface(img_prompt)
    yield final_response, image_output, map_html, debug_output_update


# ==========================================
# 3. 橋接邏輯 (Bridge Functions)
# ==========================================

def _score_to_radio_value(score):
    mapping = {1: "1 (心情差)", 2: "2 (不太好)", 3: "3 (普通)", 4: "4 (不錯)", 5: "5 (超棒)"}
    try: score = int(score)
    except: score = 3
    return mapping.get(score, "3 (普通)")

def bridge_start_click(st):
    res = emotion.on_restart(st)
    return res[1], res[0], gr.update(visible=True), res[5], res[4]

def bridge_stop_click(st):
    res = emotion.on_stop(st)
    return res[1], res[0], gr.update(visible=False), res[5], res[4]

def bridge_predict_frame(frame, st):
    res = emotion.on_stream(frame, st)
    out_cam = res[0]; out_result = res[1]; out_st = res[4]; out_btn_start = res[5]
    score_update = gr.update(); tabs_update = gr.update(); out_btn_stop = gr.update()

    if out_st.finished and hasattr(out_st, 'final_score'):
        new_val = _score_to_radio_value(out_st.final_score)
        score_update = gr.update(value=new_val)
        tabs_update = gr.Tabs(selected=1)
        out_btn_stop = gr.update(visible=False)
        
    return out_cam, out_result, out_st, out_btn_stop, out_btn_start, score_update, tabs_update


# ==========================================
# 4. Gradio 介面建構
# ==========================================

# --- [新增] 自動適應深色/淺色模式的 CSS ---
# 這段 CSS 會偵測瀏覽器設定，強制將 Markdown 區塊 (.prose) 的文字顏色反轉
adaptive_css = """
@media (prefers-color-scheme: dark) {
    /* 深色模式：文字變淺灰/白 */
    .prose, .prose > p, .prose > ul > li, .prose > ol > li, .prose span {
        color: #e2e8f0 !important; 
    }
    /* 標題變全白 */
    .prose h1, .prose h2, .prose h3, .prose h4 {
        color: #f8fafc !important;
    }
    /* 粗體字變亮 */
    .prose strong {
        color: #fbbf24 !important; /* 稍微帶一點金黃色讓重點突出，或改 #ffffff */
    }
    /* 表格文字 */
    .prose table, .prose th, .prose td {
        color: #e2e8f0 !important;
        border-color: #475569 !important;
    }
    /* Label 標籤文字 (如 "AI 回應", "心情日記") */
    span.svelte-1gfkn6j, label.svelte-1gfkn6j, .block-label {
        color: #e2e8f0 !important;
    }
    /* 除錯 Markdown 區塊 */
    .markdown-body {
        color: #e2e8f0 !important;
    }
}
@media (prefers-color-scheme: light) {
    /* 淺色模式：確保文字是深色 (防止預設主題干擾) */
    .prose, .prose > p, .prose > ul > li, .prose > ol > li {
        color: #1f2937 !important;
    }
    .prose h1, .prose h2, .prose h3 {
        color: #111827 !important;
    }
}
"""

# 將原本的 emotion.css 與新的適應性 CSS 合併
combined_css = emotion.css + adaptive_css

with gr.Blocks(title="AI 心情食堂", css=combined_css) as demo:
    
    st_state = gr.State(emotion.AppState())

    with gr.Tabs() as tabs:
        
        # Tab 1: 情緒辨識
        with gr.TabItem("😊 情緒辨識 (Step 1)", id=0):
            with gr.Column(elem_id="app_container"):
                gr.Markdown("### 第一步：測測你的心情能量\n讓 AI 看看你的表情，自動幫你決定心情分數！(辨識完畢會自動跳轉)")
                
                with gr.Row():
                    btn_start = gr.Button("📸 開啟攝影機辨識", variant="primary")
                    btn_stop  = gr.Button("⏹️ 停止", variant="secondary", visible=False)

                cam = gr.Image(sources=["webcam"], streaming=True, type="numpy", label="攝影機畫面", visible=False)
                result_markdown = gr.Markdown(emotion._hint_html("請按「開啟攝影機辨識」並允許瀏覽器使用相機。"))

        # Tab 2: 主功能區
        with gr.TabItem("🍽️ AI 心情食堂 (Step 2)", id=1):
            with gr.Column():
                gr.Markdown(f"## 🍱 今天想吃點什麼？")
                with gr.Row():
                    with gr.Column(scale=1):
                        score_input = gr.Radio(
                            ["1 (心情差)", "2 (不太好)", "3 (普通)", "4 (不錯)", "5 (超棒)"], 
                            label="1. 心情分數 (由 Tab 1 自動填入)", 
                            value="3 (普通)"
                        )
                        food_input = gr.Radio(["吃飯", "吃麵", "隨便"], label="2. 想吃什麼", value="隨便")
                        diary_input = gr.Textbox(lines=4, label="3. 心情日記", placeholder="寫下今天發生的事...")
                        debug_mode_btn = gr.Checkbox(label="🔧 開啟除錯模式", value=False)
                        
                        submit_btn = gr.Button("🍱 送出給 Agent", variant="primary")
                        
                        # --- [新增] 除錯資訊顯示區 (位於按鈕下方) ---
                        debug_output = gr.Markdown(label="除錯資訊 (Debug Log)", visible=False)
                        # ----------------------------------------
                    
                    with gr.Column(scale=1):
                        agent_output = gr.Markdown(label="AI 回應")
                        image_output = gr.Image(label="AI 推薦美食圖", type="pil", width=400)
                        map_output = gr.HTML(label="地圖導航")

    # ==========================================
    # 事件綁定
    # ==========================================

    btn_start.click(
        fn=bridge_start_click,
        inputs=[st_state],
        outputs=[result_markdown, cam, btn_stop, btn_start, st_state],
        show_progress="minimal"
    )

    btn_stop.click(
        fn=bridge_stop_click,
        inputs=[st_state],
        outputs=[result_markdown, cam, btn_stop, btn_start, st_state],
        show_progress="minimal"
    )

    cam.stream(
        fn=bridge_predict_frame,
        inputs=[cam, st_state],
        outputs=[cam, result_markdown, st_state, btn_stop, btn_start, score_input, tabs], 
        show_progress="minimal"
    )

    submit_btn.click(
        fn=mood_agent_logic,
        inputs=[score_input, food_input, diary_input, debug_mode_btn],
        outputs=[agent_output, image_output, map_output, debug_output] # 加入 debug_output
    )

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

	import gradio as gr
	from openai import OpenAI
	from huggingface_hub import InferenceClient
	import os
	import time
	import requests
	import urllib.parse
	import pandas as pd
	from langchain_huggingface import HuggingFaceEmbeddings
	from langchain_community.vectorstores import FAISS
	from PIL import Image
	from dotenv import load_dotenv

	# --- 匯入組員的模組 ---
	import emotion
	# ---------------------------

	load_dotenv()

	# ==========================================
	# 0. 環境變數 & 1. 系統初始化 (保持不變)
	# ==========================================
	GROQ_API_KEY = os.getenv("GROQ_API_KEY")
	HF_TOKEN = os.getenv("HF_TOKEN")

	global_df = None
	global_mood_df = None
	global_retriever = None
	rag_initialized = False

	def init_rag_system():
	global global_df, global_mood_df, global_retriever, rag_initialized
	if rag_initialized: return
	try:
	global_df = pd.read_csv('restaurants.csv')
	global_df['RAG_Content'] = global_df['RAG_Content'].fillna("")
	global_df['Category'] = global_df['Category'].fillna("其他")
	global_mood_df = pd.read_csv('mood_food_guide.csv')
	except Exception: pass
	if os.path.exists("faiss_index"):
	try:
	embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
	vectorstore = FAISS.load_local("faiss_index", embeddings, allow_dangerous_deserialization=True)
	global_retriever = vectorstore.as_retriever(search_kwargs={"k": 3})
	except Exception: pass
	rag_initialized = True

	# ==========================================
	# 2. 核心功能 (修改：回傳分開的除錯資訊)
	# ==========================================
	def get_restaurant_data(mood_score_str, food_choice):
	init_rag_system()
	if global_df is None or global_df.empty: return None, True, "資料庫未載入", "無建議"
	try: score = int(str(mood_score_str).split(' ')[0])
	except: score = 3
	mood_info = global_mood_df[global_mood_df['分數'] == score]
	if not mood_info.empty:
	rec_categories = mood_info.iloc[0]['推薦料理類別']
	mood_reason = mood_info.iloc[0]['原因']
	else:
	rec_categories = ""
	mood_reason = "隨意探索"
	candidates = global_df.copy()
	if rec_categories:
	candidates = candidates[candidates['Category'].apply(lambda x: str(x) in str(rec_categories) or str(rec_categories) in str(x))]
	food_keyword = "飯" if food_choice == "吃飯" else "麵" if food_choice == "吃麵" else ""
	if food_keyword:
	candidates = candidates[candidates['Name'].str.contains(food_keyword, case=False, na=False) \| candidates['RAG_Content'].str.contains(food_keyword, case=False, na=False)]
	if candidates.empty:
	result = global_df.sample(1).iloc[0]; is_random = True
	else:
	result = candidates.sample(1).iloc[0]; is_random = False
	return result, is_random, rec_categories, mood_reason

	def generate_content_with_groq(restaurant_name, restaurant_detail, user_diary, mood_score, mood_guide_reason, debug_mode=False):
	if not GROQ_API_KEY: return "⚠️ 請設定 GROQ_API_KEY", ""

	client = OpenAI(api_key=GROQ_API_KEY, base_url="https://api.groq.com/openai/v1")

	system_prompt = "你是一個幽默、懂吃且善解人意的 AI 朋友..."
	user_msg = f"【狀態】心情分數：{mood_score}\n【日記】{user_diary}\n【建議原因】{mood_guide_reason}\n【餐廳】{restaurant_name}\n資料：{restaurant_detail}"

	# 準備除錯資訊
	debug_log = ""
	if debug_mode:
	debug_log = f"""
	### 🔧 Groq Prompt Debug
	System Prompt:
	{system_prompt}

	User Message:
	{user_msg}
	"""

	try:
	response = client.chat.completions.create(model="llama-3.3-70b-versatile", messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": user_msg}])
	content = response.choices[0].message.content
	return content, debug_log # 回傳 (內容, 除錯資訊)
	except Exception as e:
	return f"Groq Error: {str(e)}", debug_log

	def generate_image_huggingface(prompt):
	if not HF_TOKEN: return None
	try:
	hf_client = InferenceClient(token=HF_TOKEN)
	return hf_client.text_to_image(prompt=prompt, model="stabilityai/stable-diffusion-xl-base-1.0")
	except: return None

	def mood_agent_logic(score_input, food_input, diary_input, debug_mode):
	restaurant, is_random, rec_categories, mood_reason = get_restaurant_data(score_input, food_input)
	if restaurant is None: yield "資料庫讀取錯誤", None, "", gr.update(); return

	name = restaurant['Name']; address = restaurant['Address']; url = restaurant['URL']; img_prompt = restaurant.get('Visual_prompt')

	rag_info = str(restaurant.get('RAG_Content', ''))
	if global_retriever:
	docs = global_retriever.invoke(name)
	if docs: rag_info = "\n".join([d.page_content for d in docs])

	# 呼叫 Groq 生成內容與取得除錯資訊
	ai_text, groq_debug_log = generate_content_with_groq(name, rag_info, diary_input, score_input, mood_reason, debug_mode)

	# 準備完整的除錯資訊 (包含圖片 Prompt)
	full_debug_log = ""
	if debug_mode:
	img_debug_log = f"""
	### 🎨 Image Prompt Debug
	Visual Prompt:
	{img_prompt}
	"""
	full_debug_log = groq_debug_log + "\n\n" + img_debug_log

	# 設定除錯視窗的狀態 (有勾選才顯示)
	debug_output_update = gr.update(value=full_debug_log, visible=debug_mode)

	final_response = f"### 🍽️ 推薦：{name}\n\n{ai_text}"
	map_html = f'<div style="text-align:center"><a href="{url}" target="_blank" style="background:#4CAF50;color:white;padding:8px 16px;border-radius:20px;text-decoration:none">🗺️ Google Map 導航</a></div>'

	# 先回傳文字與除錯資訊
	yield final_response, None, map_html, debug_output_update

	# 再回傳圖片
	image_output = generate_image_huggingface(img_prompt)
	yield final_response, image_output, map_html, debug_output_update


	# ==========================================
	# 3. 橋接邏輯 (Bridge Functions)
	# ==========================================

	def _score_to_radio_value(score):
	mapping = {1: "1 (心情差)", 2: "2 (不太好)", 3: "3 (普通)", 4: "4 (不錯)", 5: "5 (超棒)"}
	try: score = int(score)
	except: score = 3
	return mapping.get(score, "3 (普通)")

	def bridge_start_click(st):
	res = emotion.on_restart(st)
	return res[1], res[0], gr.update(visible=True), res[5], res[4]

	def bridge_stop_click(st):
	res = emotion.on_stop(st)
	return res[1], res[0], gr.update(visible=False), res[5], res[4]

	def bridge_predict_frame(frame, st):
	res = emotion.on_stream(frame, st)
	out_cam = res[0]; out_result = res[1]; out_st = res[4]; out_btn_start = res[5]
	score_update = gr.update(); tabs_update = gr.update(); out_btn_stop = gr.update()

	if out_st.finished and hasattr(out_st, 'final_score'):
	new_val = _score_to_radio_value(out_st.final_score)
	score_update = gr.update(value=new_val)
	tabs_update = gr.Tabs(selected=1)
	out_btn_stop = gr.update(visible=False)

	return out_cam, out_result, out_st, out_btn_stop, out_btn_start, score_update, tabs_update


	# ==========================================
	# 4. Gradio 介面建構
	# ==========================================

	# --- [新增] 自動適應深色/淺色模式的 CSS ---
	# 這段 CSS 會偵測瀏覽器設定，強制將 Markdown 區塊 (.prose) 的文字顏色反轉
	adaptive_css = """
	@media (prefers-color-scheme: dark) {
	/* 深色模式：文字變淺灰/白 */
	.prose, .prose > p, .prose > ul > li, .prose > ol > li, .prose span {
	color: #e2e8f0 !important;
	}
	/* 標題變全白 */
	.prose h1, .prose h2, .prose h3, .prose h4 {
	color: #f8fafc !important;
	}
	/* 粗體字變亮 */
	.prose strong {
	color: #fbbf24 !important; /* 稍微帶一點金黃色讓重點突出，或改 #ffffff */
	}
	/* 表格文字 */
	.prose table, .prose th, .prose td {
	color: #e2e8f0 !important;
	border-color: #475569 !important;
	}
	/* Label 標籤文字 (如 "AI 回應", "心情日記") */
	span.svelte-1gfkn6j, label.svelte-1gfkn6j, .block-label {
	color: #e2e8f0 !important;
	}
	/* 除錯 Markdown 區塊 */
	.markdown-body {
	color: #e2e8f0 !important;
	}
	}

	@media (prefers-color-scheme: light) {
	/* 淺色模式：確保文字是深色 (防止預設主題干擾) */
	.prose, .prose > p, .prose > ul > li, .prose > ol > li {
	color: #1f2937 !important;
	}
	.prose h1, .prose h2, .prose h3 {
	color: #111827 !important;
	}
	}
	"""

	# 將原本的 emotion.css 與新的適應性 CSS 合併
	combined_css = emotion.css + adaptive_css

	with gr.Blocks(title="AI 心情食堂", css=combined_css) as demo:

	st_state = gr.State(emotion.AppState())

	with gr.Tabs() as tabs:

	# Tab 1: 情緒辨識
	with gr.TabItem("😊 情緒辨識 (Step 1)", id=0):
	with gr.Column(elem_id="app_container"):
	gr.Markdown("### 第一步：測測你的心情能量\n讓 AI 看看你的表情，自動幫你決定心情分數！(辨識完畢會自動跳轉)")

	with gr.Row():
	btn_start = gr.Button("📸 開啟攝影機辨識", variant="primary")
	btn_stop = gr.Button("⏹️ 停止", variant="secondary", visible=False)

	cam = gr.Image(sources=["webcam"], streaming=True, type="numpy", label="攝影機畫面", visible=False)
	result_markdown = gr.Markdown(emotion._hint_html("請按「開啟攝影機辨識」並允許瀏覽器使用相機。"))

	# Tab 2: 主功能區
	with gr.TabItem("🍽️ AI 心情食堂 (Step 2)", id=1):
	with gr.Column():
	gr.Markdown(f"## 🍱 今天想吃點什麼？")
	with gr.Row():
	with gr.Column(scale=1):
	score_input = gr.Radio(
	["1 (心情差)", "2 (不太好)", "3 (普通)", "4 (不錯)", "5 (超棒)"],
	label="1. 心情分數 (由 Tab 1 自動填入)",
	value="3 (普通)"
	)
	food_input = gr.Radio(["吃飯", "吃麵", "隨便"], label="2. 想吃什麼", value="隨便")
	diary_input = gr.Textbox(lines=4, label="3. 心情日記", placeholder="寫下今天發生的事...")
	debug_mode_btn = gr.Checkbox(label="🔧 開啟除錯模式", value=False)

	submit_btn = gr.Button("🍱 送出給 Agent", variant="primary")

	# --- [新增] 除錯資訊顯示區 (位於按鈕下方) ---
	debug_output = gr.Markdown(label="除錯資訊 (Debug Log)", visible=False)
	# ----------------------------------------

	with gr.Column(scale=1):
	agent_output = gr.Markdown(label="AI 回應")
	image_output = gr.Image(label="AI 推薦美食圖", type="pil", width=400)
	map_output = gr.HTML(label="地圖導航")

	# ==========================================
	# 事件綁定
	# ==========================================

	btn_start.click(
	fn=bridge_start_click,
	inputs=[st_state],
	outputs=[result_markdown, cam, btn_stop, btn_start, st_state],
	show_progress="minimal"
	)

	btn_stop.click(
	fn=bridge_stop_click,
	inputs=[st_state],
	outputs=[result_markdown, cam, btn_stop, btn_start, st_state],
	show_progress="minimal"
	)

	cam.stream(
	fn=bridge_predict_frame,
	inputs=[cam, st_state],
	outputs=[cam, result_markdown, st_state, btn_stop, btn_start, score_input, tabs],
	show_progress="minimal"
	)

	submit_btn.click(
	fn=mood_agent_logic,
	inputs=[score_input, food_input, diary_input, debug_mode_btn],
	outputs=[agent_output, image_output, map_output, debug_output] # 加入 debug_output
	)

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